Research Case Study — Computer Vision

Pan-Organ Tumor Detection on Whole-Slide Images

A collaborative research project with IIT Bombay's Medical Lab and Tata Memorial Centre: a generalizable computer-vision model that detects tumors across multiple human organs from gigapixel whole-slide histopathology images.

Computer VisionHistopathologyCNNMedical AI
Role
ML Researcher
Collaborators
IIT Bombay · Tata Memorial
Domain
Cancer Histopathology
Output
Research Paper

Project Overview

This research, conducted with IIT Bombay’s Medical Lab and Tata Memorial Centre, set out to build a single computer-vision model that detects tumors across multiple distinct organs from gigapixel whole-slide images (WSIs) — the ultra-high-resolution scans pathologists examine under a microscope.

It is difficult because a WSI can be billions of pixels, far too large to feed to a network directly; because staining varies between hospital labs, so a model can learn the lab instead of the biology; and because a model trained on one organ rarely generalizes to others. The goal was cross-organ generalization, not single-organ overfitting.

Why it's difficult

The Research Problem

Three constraints define the problem. First, scale: gigapixel WSIs cannot be processed whole without exhausting GPU memory, so the image must be tiled without losing diagnostically relevant regions. Second, domain shift: stain intensity differs across labs and scanners, and a model that keys on color rather than tissue structure will fail to transfer. Third, generalization: the objective was one model that works across kidney, breast, and colon tissue, not a per-organ classifier.

The research question was whether a histopathology-pretrained backbone, combined with the right preprocessing, could learn organ-agnostic tumor signal robust to those sources of variation.

My Role

What I owned on this project:

  • Led the model architecture design — selecting and fine-tuning a histopathology-pretrained CNN for the multi-organ objective.
  • Built the data-preprocessing pipeline for handling gigapixel WSI files, including patch extraction and stain normalization.
  • Worked closely with medical professionals to keep model outputs clinically interpretable and aligned with how pathologists reason.
  • Ran the cross-organ evaluation that tested whether the approach generalized beyond a single tissue type.

Organized by research problem

Technical Contributions

Multi-organ tumor classifier

Problem

Models trained on one organ rarely transfer to others; the goal was one generalizable detector.

Approach

Fine-tuned KimiaNet — a histopathology-pretrained CNN with a DenseNet-121 backbone — rather than an ImageNet model, so features started from pathology, not natural images.

Outcome

89% AUC on the multi-organ test set spanning kidney, breast, and colon tissue.

Gigapixel WSI handling

Problem

A whole-slide image is far too large to fit in GPU memory or feed to a network whole.

Approach

Engineered an adaptive patch-extraction pipeline that tiles WSIs efficiently, keeping tissue-bearing regions without exhausting VRAM.

Outcome

Tractable training and inference over gigapixel inputs on commodity GPU memory.

Cross-lab robustness

Problem

Stain intensity varies between hospital labs, letting a model learn the lab rather than the tumor.

Approach

Applied Reinhard color normalization to standardize stain appearance across sources before training.

Outcome

Cross-organ generalization that holds up despite staining variation between labs.

Reasoning, not just the result

Engineering Decisions

Why KimiaNet instead of an ImageNet-pretrained backbone?

Histopathology images look nothing like natural photographs. Starting from a pathology-pretrained network means the features already encode tissue morphology, so fine-tuning converges to relevant representations rather than relearning them from scratch.

Why patch extraction rather than downsampling the whole slide?

Downsampling a gigapixel slide to fit memory destroys the cellular detail that distinguishes tumor from healthy tissue. Tiling preserves resolution where diagnosis actually happens.

Why color normalization before anything else?

Without it, the fastest way for the model to reduce loss is to key on lab-specific stain color — a shortcut that destroys generalization. Normalizing removes that shortcut so the model has to learn tissue structure.

Why report AUC?

On imbalanced medical data, raw accuracy is misleading. AUC captures the discrimination trade-off across thresholds, which is the clinically meaningful question for a screening model.

High-level view

Pipeline

Gigapixel WSI

Whole-slide histopathology scan

Adaptive Patch Extraction

Tiles tissue regions within VRAM limits

Reinhard Color Normalization

Standardizes stain across labs

KimiaNet / DenseNet-121

Histopathology-pretrained backbone

Tumor Classification

Kidney · breast · colon

Interpretable Output

Clinically reviewable

Simplified pipeline. See the linked paper for full methodology.

Challenges & Trade-offs

Challenge

Patch size: detail vs. context.

Trade-off

Smaller patches preserve cellular detail but lose tissue-level context; larger patches capture structure but cost memory and detail. The extraction strategy had to balance both under a fixed VRAM budget.

Challenge

Normalization vs. signal loss.

Trade-off

Color normalization removes the lab-color shortcut, but overly aggressive normalization can wash out genuine biological color cues — so it had to standardize appearance without erasing diagnostic signal.

Challenge

Single generalist vs. per-organ specialists.

Trade-off

One multi-organ model generalizes and is simpler to deploy, but concedes some per-organ peak performance a dedicated specialist might reach — an intentional choice given the generalization objective.

How it was measured

Evaluation

89%AUC on multi-organ test set
3 organsKidney, breast, colon
GigapixelWSI input resolution
Cross-labStain-normalized generalization
  • AUC on a held-out multi-organ test set, chosen over accuracy because of class imbalance in medical data.
  • Cross-organ generalization tested explicitly across kidney, breast, and colon rather than a single tissue type.
  • Robustness to staining variation validated by evaluating across lab sources after normalization.
  • Clinical interpretability reviewed with medical collaborators, not just measured numerically.

Impact

  • Demonstrated that a single histopathology-pretrained model can generalize tumor detection across multiple organs.
  • Produced a reproducible preprocessing recipe (patch extraction + color normalization) for gigapixel WSIs on constrained hardware.
  • Contributed a published research artifact in collaboration with IIT Bombay and Tata Memorial Centre.

Technologies

Frameworks

PyTorch

Models

KimiaNetDenseNet-121

Methods

Reinhard NormalizationAdaptive Patch ExtractionTransfer Learning

Domain

Whole-Slide ImagesHistopathology

Lessons Learned

  • In medical AI, preprocessing (color normalization, artifact removal) often drives generalization more than raw architecture choice.
  • Pretraining domain matters: a pathology-pretrained backbone beats a natural-image one for this task despite fewer parameters spent.
  • Explainability is a first-class requirement, not a nice-to-have, when outputs inform clinical diagnosis.
  • Choosing the right metric (AUC over accuracy) is part of doing the science correctly on imbalanced data.

Public References