Multi-instance CNN for Breast Cancer Classification

Note: Project was part of a Praktikum at the Technical University of Munich. Credit goes also to Hannes Hase who worked in the project.

Problem Description

The dataset comes from the BACH breast cancer histology images challenge. The focus here is on the classification subtask, which is a multi-class classification task.

The dataset contains four classes ordered by the predominant cancer type, each class has 100 examples and images have a size of 2048 x 1536 pixels. A sample from each class is shown in the following figure.

Overview of the Solution Used

Several patches are extracted from each image and fed to a deep CNN as a bag of images. Afterwards, multi-instance learning is employed by implementing a custom pooling layer that pools over the feature dimension of the final image representation before feeding it to a classifier (e.g. a couple of feed-forward layers).

The implementation framework is adopted from Tensorflow Project Template with some customizations. The following figure shows the high-level structure of the project.

An overview of the main model architecture contains a deep CNN followed by two branches (single-instance and multi-instance), where each has its own loss, with the multi-instance branch containing the custom pooling layer. The final loss is a weighted combination of both losses. The following figure shows an overview of the model architecture. This model is adopted from the paper [1].

A couple of different CNNs were tested as backbone, with most of the results reported on the famous ResNet [2]

The custom pooling layer can perform one of three operations; average-pooling, max-pooling, or log-sum-exponent. Results for experiements with the three pooling functions are reported.

The combined weighted loss can also be varied during training, as done in [1]. The following figure shows the weights of both losses (single-instance and multi-instance).

References

1. Conjeti S., Paschali M., Katouzian A., Navab N. (2017) Deep Multiple Instance Hashing for Scalable Medical Image Retrieval. In: Descoteaux M., Maier-Hein L., Franz A., Jannin P., Collins D., Duchesne S. (eds) Medical Image Computing and Computer-Assisted Intervention − MICCAI 2017. MICCAI 2017. Lecture Notes in Computer Science, vol 10435.

2. He, K., Zhang, X., Ren, S., & Sun, J. (2016). Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition (pp. 770-778).