Task 1of three
Pneumonia detectionChest X-ray
Binary classification (pneumonia vs. normal).
- Pneumonia
- Normal
Page 1: Home — Problem Statement
Medical imaging is one of the fastest ways to spot serious disease early, but reading scans takes time, specialist training, and sustained attention. Chest X-rays, brain MRIs, and retinal fundus photographs are ordered in huge volumes, and delays in reading them can mean delayed treatment for conditions that worsen quickly.
This project asks a practical question: how well can convolutional neural networks (CNNs) classify medical images across three distinct clinical tasks—pneumonia on chest X-rays, brain tumors on MRI scans, and diabetic retinopathy on retinal fundus photos—and where do those models break down?
Rather than treating deep learning as a black box that "replaces doctors," this site walks through the full pipeline from raw pixels to clinical deployment. It covers how CNNs extract visual features, how public medical datasets are prepared, how our three models compare against published benchmarks, and why real-world deployment depends as much on handling bias, privacy, and false negatives as on test-set accuracy.
Project scope at a glance:
Task 1 — Pneumonia detection (Chest X-ray): Binary classification (pneumonia vs. normal).
Task 2 — Brain tumor classification (MRI): Four-class classification (glioma, meningioma, pituitary tumor, no tumor).
Task 3 — Diabetic retinopathy grading (Fundus photography): Severity grading and detection of referable disease.
Page 2: AI & Deep Learning in Healthcare
Artificial intelligence in healthcare is not a single technology; it is a stack of methods ranging from rule-based alerts to deep neural networks. Traditional machine learning required engineers and clinicians to hand-design features—such as lesion diameter, edge sharpness, or pixel intensity histograms—before feeding them into a classifier.
Deep learning removes that manual feature-engineering step. Given enough labeled examples, a multi-layer neural network learns its own hierarchy of features directly from raw data. In medical imaging, this matters because subtle patterns—like hazy lung opacities on an X-ray or tiny microaneurysms on a retina—are hard to define with hand-written rules.
In clinical practice today, deep learning is used primarily for three jobs:
Screening: Checking high-volume routine exams (such as diabetic eye exams) to flag patients who need a specialist.
Triage: Scanning incoming emergency imaging (such as head CTs) to push suspected critical cases to the top of a radiologist's reading queue.
Page 5: Training & Results
Dataset & Scope
Published Performance
Pneumonia CXR
Kermany et al. (2018), Cell [1]
5,232 train / 624 test pediatric chest X-rays (pneumonia vs. normal)
Accuracy 92.8%, Sensitivity 93.2%, Specificity 90.1%, AUC 96.8%
Dataset & Scope
Published Performance
Brain Tumor MRI
Wong et al. (2025), PMC [3]
Combined Figshare + SARTAJ + Br35H (4 classes: glioma, meningioma, pituitary, no tumor; VGG16 transfer learning)
Overall test accuracy 99.24%, Sensitivity 99.18%, Specificity 99.75% (per-class F1: 0.98–1.00)
Dataset & Scope
Published Performance
Diabetic Retinopathy
Gulshan et al. (2016), JAMA [2]
EyePACS-1 and Messidor-2 validation sets (referable diabetic retinopathy)
EyePACS-1 AUC 0.991 (high-spec: 90.3% sens / 98.1% spec; high-sens: 97.5% sens / 93.4% spec); Messidor-2 AUC 0.990 (high-spec: 87.0% sens / 98.5% spec; high-sens: 96.1% sens / 93.9% spec)
A key takeaway for Q&A: high internal test-set scores—even above 95%—do not guarantee clinical readiness. As Page 7 shows, models that look strong on held-out splits from the same dataset can drop sharply when tested at a new hospital [4].
Page 6: Benefits
When properly validated and scoped to a specific workflow, deep learning offers concrete advantages in medical imaging:
Faster triage for urgent findings: In emergency radiology, AI triage software can flag suspected critical scans minutes after acquisition rather than waiting in a first-in, first-out queue. In Aidoc's FDA 510(k) validation study for intracranial hemorrhage on head CT, time-to-notification was 4.5 minutes compared with a standard-of-care time-to-exam-open of 72.6 minutes—a mean difference of 68.1 minutes [10].
Expanding screening access: Autonomous screening systems like IDx-DR are designed for primary care clinics rather than specialist ophthalmology centers, allowing adults with diabetes to be screened for more-than-mild diabetic retinopathy during routine visits [8].
Consistency under fatigue: Unlike human readers, a deployed model does not tire at the end of a long shift and applies the same decision threshold to every scan.
Scalable second-read support: In high-volume settings or regions with few specialists, CNNs can help prioritize which cases need immediate expert review, letting clinicians spend more time on complex or borderline exams.
These benefits only materialize when the model is paired with clear clinical guardrails—knowing exactly which patients it was validated on, what its error rates are, and who acts on its output.
Faster triage for urgent findings: In emergency radiology, AI triage software can flag suspected critical scans minutes after acquisition rather than waiting in a first-in, first-out queue. In Aidoc's FDA 510(k) validation study for intracranial hemorrhage on head CT, time-to-notification was 4.5 minutes compared with a standard-of-care time-to-exam-open of 72.6 minutes—a mean difference of 68.1 minutes [10].
Expanding screening access: Autonomous screening systems like IDx-DR are designed for primary care clinics rather than specialist ophthalmology centers, allowing adults with diabetes to be screened for more-than-mild diabetic retinopathy during routine visits [8].
Page 9: Real-World Cases
These three cases illustrate the spectrum from peer-reviewed research to FDA-cleared triage and autonomous diagnosis.
What it does: Analyzes retinal fundus photographs taken with a Topcon NW400 camera to automatically detect more than mild diabetic retinopathy (mtmDR) in adults diagnosed with diabetes who have not previously been diagnosed with diabetic retinopathy [8]. It is designed for primary care settings and outputs a screening decision without requiring an ophthalmologist to read the image.
Regulatory status & date: Granted FDA De Novo authorization (DEN180001) on April 11, 2018, classified as a diabetic retinopathy detection device (21 CFR 886.1100, Product Code PIB) [8, 9].
Verified clinical evidence: In its pivotal study of 900 enrolled patients across 10 primary care sites (819 fully analyzable), IDx-DR achieved 87.4% observed sensitivity (95% CI: 81.9%–92.9%) and 89.5% observed specificity (95% CI: 86.9%–93.1%) [8].
Project name: dl-project.
Medical Image Classification with Deep Learning — Site Content
Build notes for Spark: Each ## Page N: section below corresponds to one page of the site in syllabus order. Model-result placeholders ([[PNEUMONIA_METRICS]], [[TUMOR_METRICS]], [[DR_METRICS]]) are left verbatim in Page 5 for injection from the trained models. Every number in this document is traced to a numbered source in Page 10.
Page 1: Home — Problem Statement
Medical imaging is one of the fastest ways to spot serious disease early, but reading scans takes time, specialist training, and sustained attention. Chest X-rays, brain MRIs, and retinal fundus photographs are ordered in huge volumes, and delays in reading them can mean delayed treatment for conditions that worsen quickly.
This project asks a practical question: how well can convolutional neural networks (CNNs) classify medical images across three distinct clinical tasks—pneumonia on chest X-rays, brain tumors on MRI scans, and diabetic retinopathy on retinal fundus photos—and where do those models break down?
Rather than treating deep learning as a black box that "replaces doctors," this site walks through the full pipeline from raw pixels to clinical deployment. It covers how CNNs extract visual features, how public medical datasets are prepared, how our three models compare against published benchmarks, and why real-world deployment depends as much on handling bias, privacy, and false negatives as on test-set accuracy.
Project scope at a glance:
Task 1 — Pneumonia detection (Chest X-ray): Binary classification (pneumonia vs. normal).
Task 2 — Brain tumor classification (MRI): Four-class classification (glioma, meningioma, pituitary tumor, no tumor).
Task 3 — Diabetic retinopathy grading (Fundus photography): Severity grading and detection of referable disease.
<!-- planning-notes:v1:start -->Planning resolved these page and owner names itself instead of stopping. Review them before building.
Named in your document, not planned
Name changes
Raw medical images cannot be fed straight into a CNN. Our preprocessing pipeline addresses four practical issues:
Resizing and aspect normalization: Standardizing variable-resolution scans to a fixed input size (for example, 224×224 pixels) expected by the backbone architecture.
Intensity normalization: Scaling pixel values to a consistent range so brightness and contrast differences across scanners do not dominate training.
Data augmentation: Applying label-preserving transformations—such as small rotations, horizontal flips (where anatomically valid), and mild brightness shifts—to reduce overfitting on limited training sets.
Patient-level splitting: Keeping all scans from the same patient strictly inside either the training, validation, or test split. If one patient's scans appear in both training and testing, the model can memorize patient-specific anatomy and inflate test scores.
Each model was trained using transfer learning from a pre-trained CNN backbone, with class-aware evaluation so high overall accuracy cannot hide poor performance on a minority class. For medical tasks, accuracy alone is not enough: we track sensitivity (recall) to see how many true cases the model catches, specificity to see how well it avoids false alarms, and F1-score / AUC to summarize the trade-off.
Our Model Results
Pneumonia Detection (Chest X-ray): [[PNEUMONIA_METRICS]]
Our Model Results
Brain Tumor Classification (MRI): [[TUMOR_METRICS]]
Our Model Results
Diabetic Retinopathy Grading (Fundus): [[DR_METRICS]]
Published Benchmarks for Reference
To sanity-check our numbers, the table below lists peer-reviewed benchmark results for the same three tasks. Direct comparisons require care because datasets, patient populations, and operating thresholds differ across studies.
Task
Reference Study
Scalable second-read support: In high-volume settings or regions with few specialists, CNNs can help prioritize which cases need immediate expert review, letting clinicians spend more time on complex or borderline exams.
What it does: Analyzes non-enhanced head CT scans to flag suspected intracranial hemorrhage (ICH) and notify radiologists so urgent cases move to the top of the worklist [10].
Regulatory status & date: Received FDA 510(k) clearance (K180647) on August 1, 2018, as Class II radiological computer-aided triage and notification software (21 CFR 892.2080, Product Code QAS) [10]. The clearance explicitly states it is for workflow triage only, not diagnosis, and does not alter the original medical image [10].
Verified clinical evidence: In a retrospective, blinded, multicenter study of 198 cases across 3 sites, BriefCase achieved 93.6% sensitivity (95% CI: 86.6%–97.6%) and 92.3% specificity (95% CI: 85.4%–96.6%), with a time-to-notification of 4.5 minutes vs. 72.6 minutes standard-of-care time-to-exam-open [10].
What it does: A two-stage deep learning system that segments 3D optical coherence tomography (OCT) retinal scans into anatomical tissue maps and then recommends one of four clinical referral categories (urgent, semi-urgent, routine, observation only) across 53 key retinal diagnoses relevant to NHS referral pathways [11].
Publication & status: Published in Nature Medicine (online August 13, 2018; September 2018 issue) as a clinical research and proof-of-concept study rather than an FDA-cleared commercial device [11].
Verified clinical evidence: Trained on 14,884 OCT scan volumes from 7,621 patients, the system achieved a 5.5% overall referral error rate on the main test set (997 patients), matching the two best retina specialists (6.7% and 6.8%) and outperforming the other six experts when reading OCT scans alone [11]. When adapted to a second OCT scanner brand, the retrained segmentation network plus unchanged classification network achieved a 3.4% error rate (4 of 116 cases) [11].
| Role | Responsibilities and goals | Inputs | Decisions / observable success | Access |
|---|---|---|---|---|
| Primary care clinician | Screen adults with diabetes for more-than-mild diabetic retinopathy during routine visits, without referral to a specialist ophthalmology center | Autonomous screening output (e.g., IDx-DR) | Screening completed within the routine visit; cases needing specialist attention identified | Autonomous screening system in the primary care clinic |
| Radiologist | Ensure suspected intracranial hemorrhage (ICH) cases reach the top of the worklist quickly; interpret images | Non-enhanced head CT scans; triage notifications from BriefCase | Urgent cases reviewed sooner; notification received well ahead of standard time-to-exam-open (4.5 min vs. 72.6 min) | Triage/notification output; original medical image unchanged (workflow triage only, not diagnosis) |
| Retina specialist / expert reader | Read 3D OCT scans and assign referral categories (urgent, semi-urgent, routine, observation only) across 53 key retinal diagnoses | 3D OCT scan volumes; two-stage segmentation and referral recommendation | Referral error rate at or below expert benchmark (system 5.5% vs. best specialists 6.7% and 6.8%) | OCT referral system output (research/proof-of-concept, not an FDA-cleared device) |
| High-volume / low-specialist clinical team | Prioritize which cases need immediate expert review so clinician time concentrates on complex or borderline exams | CNN second-read output | Cases correctly prioritized for immediate expert review | Second-read support in high-volume or specialist-scarce regions |
Supporting actors (not human personas): the autonomous screening system, the BriefCase triage-and-notification software (FDA 510(k) K180647, Class II, 21 CFR 892.2080, Product Code QAS), and the two-stage OCT deep learning system.
Persona coverage here reflects only the roles evidenced in this batch; it is not a closed or complete persona catalog.
Actors: Adult patient with diabetes; primary care clinic staff; autonomous screening system (supporting actor).
Trigger: Patient attends a routine primary care visit.
Flow:
Result: The patient is screened for more-than-mild diabetic retinopathy within the primary care setting.
Actors: Clinician or reading specialist; CNN second-read system (supporting actor).
Trigger: A case enters a high-volume reading queue or a region with few specialists.
Flow:
Result: Expert review capacity is concentrated on the cases that most need it.
Actors: Radiologist; triage-and-notification system (supporting actor).
Trigger: A non-enhanced head CT scan is acquired.
Flow:
Result: Urgent cases reach the top of the worklist; the original medical image is not altered, and the output supports workflow triage only, not diagnosis.
Evidence basis: FDA 510(k) K180647, cleared August 1, 2018 (Class II, 21 CFR 892.2080, Product Code QAS); retrospective blinded multicenter study of 198 cases across 3 sites: 93.6% sensitivity (95% CI 86.6%–97.6%), 92.3% specificity (95% CI 85.4%–96.6%), time-to-notification 4.5 minutes vs. 72.6 minutes standard-of-care time-to-exam-open.
Actors: Retina specialist or referring clinician; two-stage deep learning system (supporting actor).
Trigger: A 3D OCT retinal scan volume is available for reading.
Flow:
Result: A referral category is produced for the scan.
Evidence basis: Published in Nature Medicine (online August 13, 2018; September 2018 issue) as clinical research and proof-of-concept, not an FDA-cleared commercial device. Trained on 14,884 OCT scan volumes from 7,621 patients; 5.5% overall referral error rate on the main test set (997 patients), matching the two best retina specialists (6.7% and 6.8%) and outperforming the other six experts reading OCT scans alone. When adapted to a second OCT scanner brand, the retrained segmentation network plus unchanged classification network achieved a 3.4% error rate (4 of 116 cases).
[Design proposal; unspecified choices are defaults.]
Muse and headline. Kenya Hara's Emptiness as evidence governs the visual system for dl-project. The register is calm, careful and serious: the site shows evidence, not hype. Paper white and beige fill roughly 80% of every page, charcoal text carries 15%, and rust marks only measured values and error labels at 5%.
Colour tokens (light mode).
| Role | Token |
|---|---|
| Background | #F4F1EA |
| Surface | #FBFAF6 |
| Text | #22201D |
| Primary | #3E4A45 |
| Accent (measured values, error labels only) | #B4472F |
| Muted | #8C877C |
Typography. Headings use Zen Kaku Gothic New at Light 300, sentence case, 0.06em tracking, very large with deep margins. Body uses Shippori Mincho. The scale is 1.5 modular: 96/64/40/20/17 px on desktop and 40/30/22/17/15 px on mobile, with clamp() on every step. Inter, Roboto and system-ui are excluded.
Shape and layout. Square proportions, thin 1px rules, no rounded corners. The circle appears only for the retinal scan and the gauge. Content sits in one centred column of 9 units on a 12-unit grid; each task opens a new page block with wide top margins and a thin rule.
Imagery. Still-life scans on plain grounds — a chest X-ray, an MRI slice, a fundus photo — each desaturated inside a square frame. No stock photos of doctors.
Avoid. Gradient blobs, glass panels, rounded cards with hover lift, blue or violet accents on white, bright saturated blocks, dense dashboard tiles, bouncy motion, and the generic indigo/blue-on-white SaaS template.
Readable-text rule. Headlines, wordmarks, labels, numbers, card text and controls stay whole inside the viewport and their container at 375px, 768px and 1280px, wrapping or scaling to fit; nothing covers them. Imagery and decoration may crop, bleed, rotate or overlap as the direction asks. Scrollable or moving rows may cross the edge by design and are judged on whether each item becomes fully readable as it passes; under prefers-reduced-motion they wrap or scroll so every item can be brought fully into view.
[Design proposal; unspecified choices are defaults.]
Signature moves.
Public entry composition. A full-height paper-white field. One square retinal fundus photo sits right of centre at 38vw. The stacked headline — "How well do CNNs read medical images?" — starts at the left margin in 96px light type and crosses one third of the photo edge. The subline "Three tasks, published benchmarks, and the points where models fail." sits beneath it. A thin rust rule marks the "Read the evidence" call to action below the headline. A slow focus ring moves once across the photo. The composition recomposes only accepted content and controls; it adds no new behaviour or destination.
Evidence surfaces. The three tasks — pneumonia on chest X-ray (binary), brain tumor on MRI (four-class: glioma, meningioma, pituitary, no tumor), and diabetic retinopathy grading on fundus photos — each get their own page block. Published benchmark rows (Kermany et al. 2018; Wong et al. 2025; Gulshan et al. 2016) render as rust tabular numbers under a 1px rule, so the reader can compare internal test-set scores against the warning that scores above 95% do not guarantee clinical readiness.
[Design proposal; unspecified choices are defaults.]
Interaction Model: Static (direction) Motion Tempo: still Hero Dimensionality: flat
Landing Hero Motion Brief. Focal subject: the square retinal fundus photo at 38vw, right of centre. Input → transformation → outcome: the page loads as a composed still frame; a single slow focus ring travels once across the photo over 6 seconds and stops, drawing the eye to the scan before the headline and the rust-ruled "Read the evidence" control settle into place. Motion vocabulary: very slow fades only, 900ms, no bounce; one 6-second focus sweep per scan image; the 3D scan-plane motion runs at 20% speed. Composed first frame: paper-white field, headline crossing the photo edge, rust rule under the CTA, focus ring at rest at the photo's left edge. Reduced-motion state: the focus ring is omitted and the photo, headline, subline and CTA render as the same static composition.
3D section. The user asked for a 3D website and for guidance on which 3D animation to add. The direction resolves this as one rotating MRI slice plane on a plain beige ground at 20% speed — a single crafted object showing the product's defining state (a scan plane being read), not a scene of decorative geometry. It is built with CSS 3D transforms; no Canvas/R3F/Drei is required. Under prefers-reduced-motion the plane holds a fixed angle.
Motion guardrails. No bouncy or fast motion; no hover lift on cards. Every fade and sweep is one-shot and stops. Readable text and controls never move, crop or get covered by motion.
i have to make the 3d website so tell me which kind of 3d animation we have to add in this websites
Consistency under fatigue: Unlike human readers, a deployed model does not tire at the end of a long shift and applies the same decision threshold to every scan.
Medical images are legally protected health data, not ordinary computer-vision files. In the United States, the HIPAA Privacy Rule protects all individually identifiable health information (PHI) held by covered entities and business associates; sharing scans for AI training without patient authorization requires formal de-identification—either via Expert Determination by a qualified statistician or the Safe Harbor method, which requires removing 18 specific identifiers (including names, dates finer than year, medical record numbers, device identifiers, and full-face photos) [6]. In the European Union, GDPR Article 9 classifies data concerning health as a special category whose processing is prohibited by default unless a specific legal exception applies (such as explicit consent, medical diagnosis under professional secrecy, or public-interest research with safeguards) [7]. Stripping DICOM metadata headers and preventing re-identification from head/face MRIs are non-negotiable engineering steps before training.
No site implementation stack—framework, hosting platform, or delivery technology—is specified in the accepted requirements for dl-project. The project is a content site presenting a deep-learning medical-imaging study; the evidence names no technology for building or serving the site itself.
Defaults (explicitly assumed, not user-specified):
[[PNEUMONIA_METRICS]], [[TUMOR_METRICS]], and [[DR_METRICS]] are left verbatim on the Training & Results page for injection from the trained models; the site build must preserve them unchanged until injection.Subject-matter technology described in the content (not site stack):
Any further stack choices remain unspecified and should be confirmed before implementation.
Decision support: Providing a second read or quantitative measurement alongside a clinician's own judgment.
Crucially, regulatory bodies draw a sharp line between triage/assistive tools (which only prioritize or highlight scans for a doctor) and autonomous diagnostic tools (which output a clinical decision without a specialist reviewing the image). Most deployed imaging AI sits in the first category.
Page 4: Datasets & Preprocessing
A medical imaging model is only as reliable as the data it learns from. Across our three tasks, public benchmark datasets differ in modality, image format, class balance, and labeling method:
Chest X-rays (Pneumonia): Commonly benchmarked on pediatric chest radiographs such as the Kermany et al. dataset (5,232 training images—3,883 pneumonia and 1,349 normal—and 624 test images) [1].
Brain MRI (Tumor classification): Multi-class public collections combine sources such as Figshare, SARTAJ, and Br35H across four classes: glioma, meningioma, pituitary tumor, and no tumor [3].
Retinal Fundus Photos (Diabetic Retinopathy): Public grading benchmarks (such as EyePACS and Messidor-2) contain color photographs of the back of the eye graded by clinicians for diabetic retinopathy severity [2].
Page 7: Challenges — Wrong Predictions, Privacy, Bias & Dataset Shift
In medical imaging, errors are asymmetric. A false negative (missing a real pneumonia, brain tumor, or sight-threatening retinopathy) delays treatment and can cause permanent harm or death. A false positive (flagging a healthy scan as abnormal) causes patient anxiety, unnecessary follow-up scans, biopsies, or specialist referrals—and in a triage workflow, too many false alarms cause alert fatigue so clinicians start ignoring the tool. Choosing a model's operating threshold is therefore a clinical decision, not just a math problem: screening tools typically favor high sensitivity (fewer missed cases), while confirmatory tools favor high specificity.
CNNs are notoriously good at finding shortcuts that correlate with labels in the training set instead of learning actual pathology. In a landmark cross-hospital study of pneumonia chest X-rays across three health systems (NIH, Mount Sinai Hospital, and Indiana University), Zech et al. (2018) found that internal test performance significantly exceeded external hospital performance in 3 of 5 natural comparisons—for example, a jointly trained Mount Sinai–NIH model dropped from an internal AUC of 0.931 to an external AUC of 0.815 at Indiana University [4].
The reason was shortcut learning: CNNs could identify which hospital system a chest X-ray came from with 99.95% accuracy for NIH and 99.98% accuracy for Mount Sinai, and could distinguish inpatient vs. emergency department scans within Mount Sinai with 100% accuracy, using scanner-specific features, inverted color schemes, and metallic laterality markers [4]. Because pneumonia prevalence differed sharply across hospitals (34.2% at Mount Sinai vs. 1.2% at NIH), simply guessing based on the hospital's identity achieved an AUC of 0.861 on the combined dataset [4].
The same failure mode played out at scale during the COVID-19 pandemic. A 2021 systematic review in Nature Machine Intelligence by Roberts et al. screened 2,212 papers and evaluated 62 studies in detail; 55 of 62 had a high risk of bias in at least one domain, only 13 of 62 used external validation, and the authors concluded that none of the reviewed models were ready for clinical deployment [5].
A striking example was dataset mixing: 16 of 62 papers paired adult COVID-19 X-rays with the Kermany et al. pediatric pneumonia dataset (children aged 1–5) as controls, often without noting the age gap—meaning models could score well simply by telling children apart from adults [5]. Repackaged "Frankenstein" public datasets also introduced duplicate images across training and test splits and compressed scans out of clinical DICOM format [5].
Page 8: Future Scope
Moving medical imaging AI from student prototypes and retrospective papers to trustworthy clinical tools depends on four active directions:
External, multi-center validation: Testing models prospectively on hospitals, scanners, and patient demographics that were never seen during training, so shortcut learning is caught before deployment [4, 5].
Explainability and intermediate representations: Rather than outputting a bare probability score, architectures like the two-stage DeepMind/Moorfields pipeline produce an interpretable anatomical tissue segmentation map before making a referral decision—allowing clinicians to inspect what the model saw and making the classifier portable across different scanner brands [11].
Privacy-preserving training: Techniques such as federated learning allow hospitals to train shared models without centralizing raw patient scans, helping institutions comply with HIPAA and GDPR while pooling larger, more diverse cohorts [6, 7].
Post-deployment monitoring: Hospital scanners, software updates, and patient populations change over time. Continuous auditing of model sensitivity, specificity, and drift in live clinical use is necessary to keep performance safe after launch.
Page 10: References
Kermany, D. S., et al. (2018). Identifying Medical Diagnoses and Treatable Diseases by Image-Based Deep Learning. Cell, 172(5), 1122–1131. https://www.cell.com/fulltext/S0092-8674(18)30154-5
Gulshan, V., et al. (2016). Development and Validation of a Deep Learning Algorithm for Detection of Diabetic Retinopathy in Retinal Fundus Photographs. JAMA, 316(22), 2402–2410. https://pubmed.ncbi.nlm.nih.gov/27898976/
Wong, Y., et al. (2025). Brain tumor classification using MRI images and deep learning techniques. PMC. https://pmc.ncbi.nlm.nih.gov/articles/PMC12063847/
Zech, J. R., Badgeley, M. A., Liu, M., Costa, A. B., Titano, J. J., & Oermann, E. K. (2018). Variable generalization performance of a deep learning model to detect pneumonia in chest radiographs: A cross-sectional study. PLOS Medicine, 15(11), e1002683. https://pmc.ncbi.nlm.nih.gov/articles/PMC6219764/
Roberts, M., et al. (2021). Common pitfalls and recommendations for using machine learning to detect and prognosticate for COVID-19 using chest radiographs and CT scans. Nature Machine Intelligence, 3, 199–217. https://www.nature.com/articles/s42256-021-00307-0
U.S. Department of Health and Human Services (HHS). Summary of the HIPAA Privacy Rule (45 C.F.R. § 164.514(b)). https://www.hhs.gov/hipaa/for-professionals/privacy/laws-regulations/index.html
European Union General Data Protection Regulation (GDPR). Article 9 — Processing of special categories of personal data. https://gdpr-info.eu/art-9-gdpr/
U.S. Food and Drug Administration (FDA). De Novo Classification Request for IDx-DR (DEN180001) — Decision Summary. https://www.accessdata.fda.gov/cdrh_docs/reviews/DEN180001.pdf
U.S. Food and Drug Administration (FDA). De Novo Database Entry: IDx-DR (DEN180001). https://www.accessdata.fda.gov/scripts/cdrh/cfdocs/cfpmn/denovo.cfm?id=DEN180001
U.S. Food and Drug Administration (FDA). 510(k) Premarket Notification Summary: Aidoc BriefCase (K180647), August 1, 2018. https://www.accessdata.fda.gov/cdrh_docs/pdf18/K180647.pdf
De Fauw, J., et al. (2018). Clinically applicable deep learning for diagnosis and referral in retinal disease. Nature Medicine, 24(9), 1342–1350 (Author's version, Moorfields Eye Hospital & DeepMind). https://storage.googleapis.com/deepmind-media/DeepMind.com/Blog/a-major-milestone-for-the-treatment-of-eye-disease/De%20Fauw%20et.%20al%20Nature%20Medicine%202018.pdf
Page 3: How CNNs Read an Image
To a computer, a medical scan is just a grid of numbers representing pixel intensities—one channel for a grayscale X-ray or MRI slice, three channels (red, green, blue) for a color retinal photograph. A standard neural network flattens that grid into a long list, throwing away the spatial layout that makes an image meaningful.
A Convolutional Neural Network (CNN) preserves spatial structure by sliding small learnable filters (kernels) across the image. Each filter looks for a specific local pattern and produces a feature map showing where that pattern appears.
As an image moves deeper through the network, the features become more abstract:
Early layers detect low-level primitives—edges, contrast boundaries, and simple textures.
Middle layers combine those primitives into shapes, curves, and anatomical structures (such as rib borders, blood vessels, or ventricular outlines).
Deep layers assemble high-level patterns associated with pathology—such as lung consolidation, tumor margins, or retinal hemorrhages.
Pooling layers downsample feature maps so the network focuses on what is present rather than its exact pixel coordinate, while reducing computation.
Classification head takes the final pooled features and outputs a probability score for each class.
Because labeled medical datasets are small compared to general photo collections, most medical imaging projects use transfer learning: starting from a CNN pre-trained on millions of natural images (such as ImageNet) and fine-tuning its weights on medical scans. Even though natural photos look nothing like X-rays, the early edge and texture detectors transfer well and dramatically reduce the data needed to train an accurate model [1].

dl-project · Page 1 of 10 · Problem statement
Three tasks, published benchmarks, and the points where models fail.
Read the evidenceRather than treating deep learning as a black box that “replaces doctors,” this site walks through the full pipeline from raw pixels to clinical deployment.
How CNNs extract visual features from the pixel grid of a scan.
How public medical datasets are prepared before any model sees them.
How our three models compare against published benchmarks.
Why real-world deployment depends as much on handling bias, privacy and false negatives as on test-set accuracy.
High internal test-set scores, even above 95%, do not guarantee clinical readiness. Models that look strong on held-out splits from the same dataset can drop sharply when tested at a new hospital.
Read the evidence
dl-project · Page 1 of 10 · Problem statement
Three tasks, published benchmarks, and the points where models fail.
Read the evidenceRather than treating deep learning as a black box that “replaces doctors,” this site walks through the full pipeline from raw pixels to clinical deployment.
How CNNs extract visual features from the pixel grid of a scan.
How public medical datasets are prepared before any model sees them.
How our three models compare against published benchmarks.
Why real-world deployment depends as much on handling bias, privacy and false negatives as on test-set accuracy.
High internal test-set scores, even above 95%, do not guarantee clinical readiness. Models that look strong on held-out splits from the same dataset can drop sharply when tested at a new hospital.
Read the evidence
No comments yet. Be the first!