A Sparse Convolutional Autoencoder for Joint Feature Extraction and Clustering of Metastatic Prostate Cancer Images

Metastatic prostate cancer images contain rich and complex information about cellular features. However, due to high level pathogenomic diversity and lack of clinically-validated morphological characteristics, these images are often unlabeled or weakly labeled. Extracting meaningful features from these images using deep learning method is a challenging task. In this paper, we present an unsupervised sparse convolutional autoencoder (SCAE) model to tackle this problem. The imposed sparsity constraints enable the model to perform clustering during feature extraction. The experiments with 856 cores images show that the identified image feature clusters have distinctive cellular feature distributions. By fine-tuning the pretrained encoder module, our proposed model achieves comparable or superior performance to other benchmark models on molecular tumor subtype detection and mutation prediction, suggesting its potential for other downstream supervised tasks.