Fast and Accurate Tumor Segmentation of Histology Images using Persistent Homology and Deep Convolutional Features

Abstract: Tumor segmentation in whole-slide images of histology slides is an important step towards computer-assisted diagnosis. In this work, we propose a tumor segmentation framework based on the novel concept of persistent homology profiles (PHPs). For a given image patch, the homology profiles are derived by efficient computation of persistent homology, which is an algebraic tool from homology theory. We propose an efficient way of computing topological persistence of an image, alternative to simplicial homology. The PHPs are devised to distinguish tumor regions from their normal counterparts by modeling the atypical characteristics of tumor nuclei. We propose two variants of our method for tumor segmentation: one that targets speed without compromising accuracy and the other that targets higher accuracy. The fast version is based on the selection of exemplar image patches from a convolution neural network (CNN) and patch classification by quantifying the divergence between the PHPs of exemplars and the input image patch. Detailed comparative evaluation shows that the proposed algorithm is significantly faster than competing algorithms while achieving comparable results. The accurate version combines the PHPs and high-level CNN features and employs a multi-stage ensemble strategy for image patch labeling. Experimental results demonstrate that the combination of PHPs and CNN features outperforms competing algorithms. This study is performed on two independently collected colorectal datasets containing adenoma, adenocarcinoma, signet and healthy cases. Collectively, the accurate tumor segmentation produces the highest average patch-level F1-score, as compared with competing algorithms, on malignant and healthy cases from both the datasets. Overall the proposed framework highlights the utility of persistent homology for histopathology image analysis.

• The paper proposes a novel framework for fast and accurate tumor segmentation in histology images by combining persistent homology profiles with deep convolutional features.
• The framework includes a fast variant using exemplar-based PHP classification and an accurate variant combining PHPs with CNN features through an ensemble strategy.
• Both methods demonstrate robust, high-performance tumor segmentation on colorectal cancer datasets, highlighting the utility of persistent homology in computational pathology.

Fast and Accurate Tumor Segmentation of Histology Images using Persistent Homology and Deep Convolutional Features

The paper presents a novel approach to tumor segmentation in histological slide images, leveraging the concept of persistent homology profiles (PHPs) combined with deep convolutional neural network (CNN) features. Tumor segmentation is a critical component in computational pathology that assists in computer-aided diagnosis by automatically identifying tumor-rich areas in histology images, particularly for colorectal cancer (CRC). The proposed framework introduces two variants: a fast version prioritizing speed, and a more accurate version targeting higher segmentation accuracy.

The concept of PHPs is introduced as a means to model the atypical characteristics of tumor nuclei. PHPs are derived using an efficient computation of persistent homology—a technique from algebraic topology—to identify morphological differences between tumor and normal regions. This methodology is geared toward capturing the degree of nuclear connectivity in image patches, which is pivotal for distinguishing tumor regions from normal tissue.

Methodology

The paper proposes different strategies for tumor segmentation:

1. Fast Tumor Segmentation: This method uses PHPs to provide rapid segmentation without compromising accuracy. It achieves this by selecting exemplar patches using CNN activation maps. The exemplar-based classification is performed using symmetrized Kullback-Leibler divergence (KLD) to quantify divergence between PHPs of exemplar patches and input patches. The fast algorithm is noted for being significantly faster than competing methods.
2. Accurate Tumor Segmentation: Combining PHPs with CNN features, this variant employs a multi-stage ensemble strategy integrating topological and deep convolutional features. The comprehensive integration of PHPs enhances the classification accuracy of CNN models. Experimental results indicate that this fusion surpasses the performance of other algorithms.

Both methods demonstrate robustness across two CRC datasets, providing high precision and recall rates, and achieving superior average patch-level F1-scores compared to competing algorithms.

Results

Experimentation on datasets from different institutions confirms the viability and robustness of the proposed framework, particularly in handling stain variability and morphological characteristics across different centers. Furthermore, the embodied utility of persistent homology in histopathology image analysis is emphasized.

The fast tumor segmentation variant is highlighted for its computational efficiency, handling large datasets effectively with significantly lower processing times than competing algorithms. Accurate tumor segmentation showcases improved performance with the synergy of PHPs and CNN features, facilitating better segmentation accuracy on complex epithelial tumor cases.

Implications and Future Directions

The presented framework contributes to the field of computational pathology by providing an efficient and reliable method for tumor segmentation in histological images, addressing key challenges in manual segmentation processes. The PHP methodology offers potential for adaptation in other histopathological tasks, emphasizing its robustness and biological interpretability.

Looking forward, enhancements could involve refining exemplar selection processes and exploring recurrent networks to leverage temporal information in PHPs, broadening application scopes beyond current limitations. Future work may also incorporate PHPs into advancing machine learning models for histopathological image analysis, driving innovations in automated cancer diagnostics.