[Paper Review] Inception Architecture and Residual Connections in Classification of Breast Cancer Histology Images

Mohammad Ibrahim Sarker, Hyongsuk Kim+2 more|arXiv (Cornell University)|Dec 10, 2019

AI in cancer detection17 references4 citations

TL;DR

This paper applies the Inception v4 architecture with residual connections to classify breast cancer histology images from the ICIAR-2018 challenge, using extensive data augmentation and ensemble learning. It achieves 89% accuracy on a 4-class classification task and 93.7% on a binary carcinoma/non-carcinoma task, securing 23rd place among 50 teams.

ABSTRACT

This paper presents results of applying Inception v4 deep convolutional neural network to ICIAR-2018 Breast Cancer Classification Grand Challenge, part a. The Challenge task is to classify breast cancer biopsy results, presented in form of hematoxylin and eosin stained images. Breast cancer classification is of primary interest to the medical practitioners and thus binary classification of breast cancer images have been under investigation by many researchers, but multi-class categorization of histology breast images have been challenging due to the subtle differences among the categories. In this work extensive data augmentation is conducted to reduce overfitting and effectiveness of committee of several Inception v4 networks is studied. We report 89% accuracy on 4 class classification task and 93.7% on carcinoma/non-carcinoma two class classification task using our test set of 80 images.

Motivation & Objective

• To address the challenge of multi-class classification in breast cancer histology images, which exhibit subtle morphological differences across categories.
• To improve generalization in deep learning models by applying extensive data augmentation to reduce overfitting.
• To evaluate the effectiveness of ensemble learning using multiple Inception v4 networks for improved classification performance.
• To achieve high accuracy in both binary (carcinoma/non-carcinoma) and multi-class (four-category) classification of H&E stained breast biopsy images.

Proposed method

• The study employs the Inception v4 architecture, which uses inception modules with parallel convolutional filters of multiple sizes to capture multi-scale features.
• Residual connections are incorporated into the Inception v4 network to ease training of deep architectures and improve gradient flow.
• Extensive data augmentation is applied, including random cropping, flipping, rotation, and color jittering, to increase training data diversity and reduce overfitting.
• An ensemble of multiple trained Inception v4 models is used to improve prediction robustness and accuracy through voting or averaging.
• The model is trained and evaluated on the ICIAR-2018 Breast Cancer Classification Grand Challenge dataset, using a test set of 80 images.
• Evaluation is performed using standard metrics including accuracy, with results reported on both 4-class and binary classification tasks.

Experimental results

Research questions

• RQ1 Can the Inception v4 architecture with residual connections effectively classify subtle histological differences in breast cancer types?
• RQ2 How does extensive data augmentation impact model generalization and performance on limited medical imaging datasets?
• RQ3 Does ensemble learning with multiple Inception v4 models improve classification accuracy over single models?
• RQ4 What level of performance can be achieved on both binary and multi-class breast cancer histology image classification tasks?

Key findings

• The model achieved 89% accuracy on the 4-class classification task, demonstrating strong performance on fine-grained histological distinctions.
• For the binary carcinoma/non-carcinoma classification, the model reached 93.7% accuracy on the test set of 80 images.
• The use of extensive data augmentation significantly reduced overfitting, improving model generalization on the validation and test sets.
• The ensemble of multiple Inception v4 networks outperformed single models, indicating improved robustness and predictive stability.
• The approach secured 23rd place out of 50 accepted teams in the ICIAR-2018 Breast Cancer Classification Grand Challenge.