[Paper Review] Convolution Neural Networks for diagnosing colon and lung cancer histopathological images

Sanidhya Mangal, Aanchal Chaurasia+1 more|arXiv (Cornell University)|Sep 8, 2020

AI in cancer detection33 references86 citations

TL;DR

A shallow CNN is trained on the LC25000 histopathology dataset to classify lung and colon cancer images, achieving around 98% accuracy on both tasks without heavy preprocessing. The study analyzes architecture, training strategy, and results.

ABSTRACT

Lung and Colon cancer are one of the leading causes of mortality and morbidity in adults. Histopathological diagnosis is one of the key components to discern cancer type. The aim of the present research is to propose a computer aided diagnosis system for diagnosing squamous cell carcinomas and adenocarcinomas of lung as well as adenocarcinomas of colon using convolutional neural networks by evaluating the digital pathology images for these cancers. Hereby, rendering artificial intelligence as useful technology in the near future. A total of 2500 digital images were acquired from LC25000 dataset containing 5000 images for each class. A shallow neural network architecture was used classify the histopathological slides into squamous cell carcinomas, adenocarcinomas and benign for the lung. Similar model was used to classify adenocarcinomas and benign for colon. The diagnostic accuracy of more than 97% and 96% was recorded for lung and colon respectively.

Motivation & Objective

• Motivate computer-aided diagnosis of lung and colon cancer from histopathology images.
• Evaluate whether a shallow CNN can effectively classify squamous cell and adenocarcinomas in lung, and adenocarcinomas in colon.
• Show that high-resolution histopathology images can be learned without excessive downsampling.
• Provide a training/evaluation protocol and release model artifacts for replication.

Proposed method

• Use LC25000 dataset with 5,000 images per class across five classes, expanded to 25,000 via augmentation.
• Construct a three-convolution-layer CNN (32, 64, 64 filters) with 3x3 kernels and stride 2, followed by two fully connected layers (512 and 3 or 2 outputs).
• Apply ReLU activations, max-pooling after each conv layer, and dropout between dense layers.
• Train with RMSprop, mini-batch size 32, learning rate 1e-4, rho 0.9, epsilon 1e-7, and categorical cross-entropy loss for 100 iterations.
• Evaluate on an 80-10-10 train/validation/test split, reporting image-level accuracy and loss; implement the model in TensorFlow and share h5 artifacts.

Experimental results

Research questions

• RQ1 Can a shallow CNN accurately classify lung histopathology into adenocarcinoma, squamous cell carcinoma, or benign?
• RQ2 Can the same architecture classify colon histopathology into adenocarcinoma or benign with high accuracy?
• RQ3 Does a shallow CNN trained on LC25000 outperform traditional ML methods and transfer-learning approaches on the same data?

Key findings

• Lung model achieved training accuracy 97.9216% with loss 5.83230; validation accuracy 97.8987% with loss 6.11450%.
• Colon model achieved training accuracy 96.9503% with loss 7.9340%; validation accuracy 96.6110% with loss 9.7141%.
• Both models converged around 20 epochs; dropout introduced variability in accuracy/loss plots but aided generalization.
• The approach demonstrates that a shallow CNN can effectively classify high-resolution histopathology images without heavy downsampling.
• Feature maps indicate learned filters resemble texture-like patterns similar to Gabor filters, corroborating learned representations.