Towards More Transparent and Accurate Cancer Diagnosis with an Unsupervised CAE Approach

TL;DR

This work addresses diagnostic efficiency and accuracy in cancer pathology by proposing an unsupervised Content-Based Medical Image Retrieval (CBMIR) system that uses a customized Convolutional Autoencoder (CAE) to learn 200-dimensional feature embeddings from WSIs. The CAE is trained with skip connections and an attention-enabled bottleneck, and embeddings are indexed for fast patch-based retrieval using Euclidean distance, enabling top-$K$ similar patches to be shown to pathologists. Evaluations on BreaKHis, SICAPv2, and Arvaniti demonstrate high top-$5$ recall and strong precision, with competitive performance against supervised baselines and demonstrated generalization to external data. The approach aims to reduce pathologist workload, improve diagnostic confidence, and extend CBMIR applicability across cancer types and imaging modalities, potentially accelerating and improving cancer diagnosis in diverse clinical settings.

Abstract

Digital pathology has revolutionized cancer diagnosis by leveraging Content-Based Medical Image Retrieval (CBMIR) for analyzing histopathological Whole Slide Images (WSIs). CBMIR enables searching for similar content, enhancing diagnostic reliability and accuracy. In 2020, breast and prostate cancer constituted 11.7% and 14.1% of cases, respectively, as reported by the Global Cancer Observatory (GCO). The proposed Unsupervised CBMIR (UCBMIR) replicates the traditional cancer diagnosis workflow, offering a dependable method to support pathologists in WSI-based diagnostic conclusions. This approach alleviates pathologists' workload, potentially enhancing diagnostic efficiency. To address the challenge of the lack of labeled histopathological images in CBMIR, a customized unsupervised Convolutional Auto Encoder (CAE) was developed, extracting 200 features per image for the search engine component. UCBMIR was evaluated using widely-used numerical techniques in CBMIR, alongside visual evaluation and comparison with a classifier. The validation involved three distinct datasets, with an external evaluation demonstrating its effectiveness. UCBMIR outperformed previous studies, achieving a top 5 recall of 99% and 80% on BreaKHis and SICAPv2, respectively, using the first evaluation technique. Precision rates of 91% and 70% were achieved for BreaKHis and SICAPv2, respectively, using the second evaluation technique. Furthermore, UCBMIR demonstrated the capability to identify various patterns in patches, achieving an 81% accuracy in the top 5 when tested on an external image from Arvaniti.

Figures (9)

• Figure 1: proposed CAE architecture with kernel size of $3$ throughout the model, the stride of $2$ in the encoder and decoder, and $1$ in the bottleneck layer.
• Figure 2: an overview of the UCBMIR for retrieving similar cases to a given query. The preprocessing stage involves extracting tissue from the patient's body and dividing the whole slide images (WSIs) into patches of a specific sizegoing. In the training stage, these patches are used to train the proposed unsupervised CAE to extract feature representations. The trained encoder and bottleneck layers are then used to extract feature embeddings FEs that are used in the CBMIR section. In the CBMIR stage, the search engine computes the embedding features of the training set and stores them in a dictionary. When a query image is selected from the test set, the FE computes the embedding of that query and compares it with those in the dictionary. The model then returns the K most similar patches based on the pathologists' needs.
• Figure 3: the structure of training VGG as a multi-class classification on SICAPv2 and delivering $200$ features per image to the following CBMIR steps. The Conv2D layers are shown in "orange", MaxPooling2D in "red", Dense layers in "green", and Flatten in "teal".
• Figure 4: evaluation of the UCBMIR at k = 5 on BreaKHis \ref{['fig:pie chart.']} and SICAPv2 \ref{['fig:bar chart.']}. From 2122 query images in SICAPv2, for 447 cases, the model could not find at least one correct similar images according to their labels while it retrieved two similar images at 5 top for 641 cases.
• Figure 5: confusion matrix of UCBMIR on the different test cohorts at K = 3, 5, 7. a. ARVANITI (Panda), b. SICAPv2, c. ARAVNITI (SICAP).