[Paper Review] A Study of Deep Learning Colon Cancer Detection in Limited Data Access Scenarios

Apostolia Tsirikoglou, Karin Stacke+3 more|arXiv (Cornell University)|May 20, 2020

AI in cancer detection3 references4 citations

TL;DR

This study investigates transfer learning strategies for deep learning-based metastatic colon cancer detection in lymph nodes with limited annotated data. By leveraging primary colon tumor tissue or breast cancer data (with or without Cycle-GAN domain adaptation), the model achieves performance comparable to training on full lymph node datasets, demonstrating that cross-tissue and cross-cancer data can effectively mitigate data scarcity in histopathology AI.

ABSTRACT

Digitization of histopathology slides has led to several advances, from easy data sharing and collaborations to the development of digital diagnostic tools. Deep learning (DL) methods for classification and detection have shown great potential, but often require large amounts of training data that are hard to collect, and annotate. For many cancer types, the scarceness of data creates barriers for training DL models. One such scenario relates to detecting tumor metastasis in lymph node tissue, where the low ratio of tumor to non-tumor cells makes the diagnostic task hard and time-consuming. DL-based tools can allow faster diagnosis, with potentially increased quality. Unfortunately, due to the sparsity of tumor cells, annotating this type of data demands a high level of effort from pathologists. Using weak annotations from slide-level images have shown great potential, but demand access to a substantial amount of data as well. In this study, we investigate mitigation strategies for limited data access scenarios. Particularly, we address whether it is possible to exploit mutual structure between tissues to develop general techniques, wherein data from one type of cancer in a particular tissue could have diagnostic value for other cancers in other tissues. Our case is exemplified by a DL model for metastatic colon cancer detection in lymph nodes. Could such a model be trained with little or even no lymph node data? As alternative data sources, we investigate 1) tumor cells taken from the primary colon tumor tissue, and 2) cancer data from a different organ (breast), either as is or transformed to the target domain (colon) using Cycle-GANs. We show that the suggested approaches make it possible to detect cancer metastasis with no or very little lymph node data, opening up for the possibility that existing, annotated histopathology data could generalize to other domains.

Motivation & Objective

• Address the challenge of limited annotated lymph node data for metastatic colon cancer detection.
• Investigate whether primary colon tumor tissue or cross-organ cancer data can serve as effective data sources when lymph node data are scarce.
• Evaluate the impact of domain adaptation using Cycle-GANs to align breast cancer histology to colon cancer appearance.
• Assess whether transfer learning from non-lymph node sources can improve model performance in low-data scenarios.
• Explore practical data strategies that reduce annotation burden while maintaining diagnostic performance in digital pathology.

Proposed method

• Trained a 3-convolutional-layer CNN with a fully connected layer for patch-level classification.
• Used patient-level train/test splits (32/5 patients) to ensure data independence across folds.
• Evaluated three annotation effort levels: high (full lymph node data), low (primary tumor only), and medium (primary tumor + few lymph node patients).
• Augmented training data with patches from the CAMELYON17 breast cancer dataset (50 whole-slide images, 43 patients).
• Applied Cycle-GANs to transform CAMELYON17 patches into a colon-like appearance to improve domain alignment.
• Balanced all datasets at the patch level to ensure equal class distribution between tumor and non-tumor samples.

Experimental results

Research questions

• RQ1 Can primary colon tumor tissue be used as a proxy for training metastatic colon cancer detection in lymph nodes when lymph node data are limited?
• RQ2 Does transferring knowledge from a different cancer type (e.g., breast cancer) improve performance in low-data lymph node detection?
• RQ3 Does domain adaptation via Cycle-GANs enhance performance when transferring from breast to colon cancer histology?
• RQ4 How does performance vary across different levels of annotation effort (high, medium, low) in lymph node data scarcity?
• RQ5 Can data augmentation with external cancer datasets outperform training on a full but limited local lymph node dataset?

Key findings

• Augmenting a small lymph node dataset with primary colon tumor patches significantly improved model performance, even surpassing the baseline trained on the full local LNCO dataset in the medium annotation effort scenario.
• Adding CAMELYON17 breast cancer data to the training set consistently improved performance across all annotation effort levels, with the medium effort setup exceeding the performance of the full LNCO-only baseline.
• Cycle-GAN-based domain adaptation from breast to colon histology did not improve performance and may have degraded it, likely due to limited training data for the GAN.
• The model achieved high patch-level accuracy using only 25% of the primary tumor data from LNCO and data from just two lymph node patients in the medium effort setting.
• The results demonstrate that cross-tissue and cross-cancer data transfer is viable and effective in low-data scenarios for colon cancer metastasis detection.
• The study shows that existing annotated histopathology data from other cancer types can be leveraged to train high-performing models even without access to large amounts of target-domain (lymph node) data.