Source-linked AI summary
Fast and Accurate Tumor Segmentation of Histology Images using Persistent Homology and Deep Convolutional Features
Talha Qaiser, Yee-Wah Tsang, Daiki Taniyama, Naoya Sakamoto, Kazuaki Nakane, David Epstein, Nasir Rajpoot
TL;DR
Tumor segmentation supports computer-assisted diagnosis, while existing approaches have limited scope and CNNs can struggle to learn relevant components. The paper presents an automated framework using persistent homology and deep convolutional features, with experiments demonstrating efficacy and improved performance across independent pathology centers.
Problem
Tumor segmentation is a vital step toward computer-assisted diagnosis, but existing algorithms are mainly limited to hand-picked visual components and CNNs have difficulty learning some components.
Method
The framework presents two persistent homology methods and uses a simple, rapid computational procedure for extracting components.
Results
The combination of topological features and deep convolutional features demonstrates improved performance, with experiments from two independent pathology centers demonstrating efficacy and generalizability.
Takeaways & Limitations
The presented automated tumor segmentation framework supports the utility of persistent homology for histopathology image analysis.
Takeaways & Limitations
The work is limited by parameters such as the number of exemplar patches, while related algorithms are mainly limited to hand-picked visual components.
Abstract
from arXiv · showhide
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.
1. Introduction
The paper motivates automated tumor localization in colorectal histology WSIs and introduces persistent-homology-based methods targeting speed and accuracy.
- Motivation: Manual tumor-region segmentation is challenging and time-consuming, motivating automated localization for computer-assisted diagnosis.Automated methods can provide reproducible results and potentially reduce inter-observer variability.
- Motivation: Accurate tumor-rich-region segmentation may support disease-aggressiveness assessment, high-power-field selection, grading, and gene-expression profiling.Precise localization can reduce the association of nonmalignant stroma with gene-expression measurements.
- Biological basis: Tumor nuclei exhibit atypical size, chromatin texture, shape, and clustering, whereas nuclei in normal regions retain more typical morphology.These morphological differences motivate modeling nuclear structure to distinguish tumor from normal tissue.
- Contributions: Persistent homology is proposed to measure morphological differences between normal and cancer nuclei through persistent homology profiles.The framework introduces fast and accurate variants for tumor segmentation in H&E-stained WSIs.
- Results: The fast algorithm is significantly faster than conventional CNN and competing approaches while remaining competitive in accuracy.The accurate method’s combination of topological and CNN features produces high segmentation accuracy on both datasets.
- Proposed methods: The fast method compares input-patch PHPs with exemplar-patch PHPs using symmetrized Kullback-Leibler divergence and a k-NN classifier.CNN activations select exemplars, while the CNN itself is not reused after precomputation.
- Proposed methods: The accurate method combines PHPs with deep convolutional features in a multi-stage ensemble strategy for patch labeling.Experiments use malignant and healthy cases from two independent institutions and include weak and strong cross-validation.
2. Related Work
Prior histology tumor-segmentation methods use handcrafted, weakly supervised, or deep features, but limitations include feature-selection burden, restricted scope, and computational cost.
- Handcrafted features: Existing methods include handcrafted texture, color, morphological, Gabor, histogram, superpixel, and local-binary-pattern features.Other approaches use color graphs, multiscale transforms, or related visual descriptors.
- Weak supervision: Weakly supervised multiple-instance approaches generate bags of selected patches and learn from general color, texture, and transform-based features.These methods have been applied to colon-cancer tissue microarrays.
- Limitations of prior methods: Selecting optimal features for supervised or weakly supervised learning is onerous and can over-emphasize dataset-specific characteristics.This creates a risk that the learned representation is overly tied to a particular dataset.
- Limitations of prior methods: Many prior algorithms are mainly limited to hand-picked visual fields or tissue microarrays rather than whole-slide images.A clinical solution should scale segmentation results to the WSI level.
- Deep learning: Deep learning learns hierarchical data-driven features and has achieved strong computer-vision and medical-image-processing performance.U-Net-style architectures combine downsampling and upsampling convolution layers for segmentation.
- Deep learning: Deep feature-based segmentation can be computationally expensive, encounter vanishing gradients, and struggle to learn rotationally invariant characteristics.The paper contrasts these issues with PHPs that capture nuclear connectedness and are rotationally invariant.
3. Introduction to Persistent Homology
The paper defines persistent homology through image filtrations and derives compact PHPs by tracking connected components and holes across intensity thresholds.
- Persistent homology: Persistent homology studies the homology of a filtered space, represented as a nested sequence of subspaces from the empty set to the full space.The sequence is called a filtration of the topological space.
- Image filtration: For a 2D grayscale image, each filtration element is the union of closed pixels below a threshold, with the final element equal to the entire image.Threshold levels binarize the image and produce one filtration for each grayscale image.
- Image filtration: The method tracks Betti numbers β0 and β1, representing connected components and one-dimensional voids across filtration levels.Only homology dimensions 0 and 1 are needed for the described image analysis.
- Efficient computation: Careful threshold selection balances retaining grayscale information against sparsity and computation speed.Restricting computation to well-chosen thresholds can substantially improve speed with negligible information loss.
- Persistent homology profiles: The resulting PHPs record homology-group ranks across thresholds and provide an alternative to simplicial homology for 2D images.The algebraic invariants reduce to whole-number ranks, supporting the fast tumor-segmentation algorithm.
4. The Proposed Approaches
The proposed framework uses persistent homology profiles to characterize tumor morphology and offers fast and accurate segmentation variants. The accurate variant combines PHPs with deep convolutional features, while the fast variant classifies patches using representative exemplars and PHP divergence.
- Overview: The framework presents two PHP-based approaches: fast tumor segmentation and a higher-accuracy variant.Both approaches divide whole-slide images into patches and label each as tumor or non-tumor.
- Fast Tumor Segmentation: The fast method selects representative tumor and non-tumor patches from CNN activation maps and classifies inputs using PHP divergence.It computes symmetric KLD distances between an input patch and exemplar PHPs, then derives similarity from distances to nearest exemplars.
- Persistent Homology Profiles: Tumor and non-tumor regions exhibit different homology invariants, reflected in their PHP trends across thresholds.Tumor regions show slower changes in β0 and β1 during merging and formation of new classes than non-tumor regions.
- Selection of Exemplar Patches: Exemplar selection uses class-specific activation distributions and interquartile-range bins rather than only the most highly activated patches.This design addresses the risk of selecting patches representing only one tumor or normal-tissue type and differs from the cited prior approach.
- Accurate Tumor Segmentation: The accurate method combines deep convolutional features with persistent homology features in a multi-stage ensemble strategy.The authors describe this combination as using the strengths of both feature types for tumor segmentation.
5.1. Dataset and Experimental Setup
The experiments use two colorectal whole-slide image datasets with expert annotations and balanced tumor/non-tumor patch sampling. Training and testing use augmented 256 × 256 patches and CNN optimization with Adam, dropout, and Xavier initialization.
- Datasets: The Warwick-UHCW dataset contains 75 H&E-stained whole-slide images of colorectal tissue.The supplied passages also state that the slides contain more than 10^10 pixels at the highest resolution.
- Patch Sampling: For Warwick-UHCW, 75,000 patches were extracted for training and 37,500 for testing from 50 and 25 whole-slide images, respectively.Each whole-slide image contributed 1,500 randomly selected patches, including 750 tumor and 750 non-tumor patches.
- Datasets: The Warwick-Osaka dataset contains 50 H&E-stained colorectal whole-slide images scanned at 0.23µm/pixel using a Hamamatsu NanoZoomer 2.0-HT.Ground truth was handmarked by two expert pathologists, and cases span adenoma, differentiated adenocarcinoma, healthy, and signet categories.
- Patch Sampling: The Warwick-Osaka sampling selected 1,500 patches per whole-slide image, with 750 tumor and 750 non-tumor patches.The dataset was used to evaluate robustness, according to the supplied experimental description.
- Experimental Setting: Experiments split whole-slide images into 256 × 256 patches and applied rotations, flips, and hue perturbations for data augmentation.The augmentation was applied to both training datasets to counter overfitting.
5.2. Comparative Analysis
The comparative analysis evaluates exemplar-selection strategies and tumor-segmentation methods on colorectal histology data. The PHP-based fast method emphasizes computational efficiency, while the PHP+CNN accurate method achieves strong segmentation performance across datasets and cases.
- Exemplar selection: The experiment compares random, k-means, CNN-activation, and proposed exemplar-selection strategies on 75 colorectal adenocarcinoma WSIs.Random selection was repeated 10 times, and equal numbers of exemplars were used for comparison.
- Patch-level comparison: Table 1 reports mean precision, recall, and F1-score for the compared patch-based tumor-segmentation approaches.Overall results favor the proposed method, with representative tumor and non-tumor patches shown in Figure 6.
- Adenocarcinoma comparison: The PHP+CNN accurate method performs best for recall and F1-score, while PHP-based fast segmentation performs best for precision.The PHP+CNN method outperforms competing methods by a reasonable margin.
- Topological features: PHP captures rotational invariance and distinguishes tumor from non-tumor nuclei through differences in their connectivity.The PHP remains persistent under patch rotation and models a biological difference between the classes.
- Ensemble strategy: The accurate method combines topological and deep features through an ensemble strategy, whose agreement improves identification of non-tumor regions.Specificity analysis reports that agreement between topological and deep features outperforms competing methods.
- Computational efficiency: The fast segmentation algorithm is approximately 4.2 times faster than CNN and 5.2 times faster than the accurate method.The method is described as an order of magnitude faster than competing algorithms, presenting an accuracy-efficiency trade-off.
6. Discussion and Conclusions
The study presents an automated persistent-homology framework for colorectal histology whole-slide tumor segmentation, validated across two independent datasets. Its fast and accurate variants offer a speed–accuracy trade-off, while combining topological and deep convolutional features improves segmentation performance across tumor grades and tissue variability.
- Results and limitations: Combining topological and deep convolutional features improved overall CNN performance for accurate tumor segmentation across the evaluated histology cases.The discussion reports better performance despite stain and morphological variability, while benign epithelial and lymphocytic regions remained difficult for nearly all algorithms.
- Results and limitations: The method’s parameters, including the number of exemplar patches and k in k-means, were selected empirically and may require data-dependent fine-tuning.The authors identify parameter tuning as a limitation of the work.
- Fast segmentation: The fast variant uses CNN-learned activation maps to select representative exemplar patches, reducing computational expense and improving robustness to outliers.Careful exemplar selection can nearly eliminate the need to retrain on a new dataset.
- Accurate segmentation: The accurate variant combines topological signatures with deep convolutional features in a multi-stage classification strategy.The study describes this combination as a simple way to integrate complementary topological and convolutional information.
- Framework and contribution: The framework introduces persistent homology profiles as topological signatures for automated colorectal cancer histology tumor segmentation.The profiles are presented as an interpretable representation of cell connectivity and as resembling clinicians’ identification of tumor-enriched areas.
- Validation: The framework was validated on two independent datasets containing malignant and healthy cases, with experiments conducted on challenging colorectal histology data.The evaluation covered colorectal epithelial tumors of different histology grades and variation in stain and morphology.