Source-linked AI summary
Y-Net: Joint Segmentation and Classification for Diagnosis of Breast Biopsy Images
Sachin Mehta, Ezgi Mercan, Jamen Bartlett, Donald Weave, Joann G. Elmore, Linda Shapiro
TL;DR
Breast biopsy diagnosis needs both tissue-level structure and identification of diagnostically important regions, while existing segmentation and classification methods provide only part of this information. Y-Net jointly generates tissue segmentation and discriminative maps through a modular U-Net extension, achieving state-of-the-art segmentation with fewer parameters and higher diagnostic classification accuracy.
Problem
Existing segmentation methods lack sufficient labeled data and cannot weight tissue types by diagnostic relevance, while classification methods lack structure- and tissue-level information.
Method
Y-Net extends U-Net with a parallel discriminative-map branch and modular convolutional blocks, combining tissue masks and diagnostic maps for classification.
Results
Y-Net matches state-of-the-art segmentation accuracy with 6.6× fewer parameters and improves diagnostic classification accuracy by about 9% using discriminative instead of segmentation masks.
Takeaways & Limitations
Discriminative tissue-level segmentation masks provide powerful features for breast biopsy diagnosis and support higher accuracy than state-of-the-art segmentation- or saliency-based methods.
Abstract
from arXiv · showhide
In this paper, we introduce a conceptually simple network for generating discriminative tissue-level segmentation masks for the purpose of breast cancer diagnosis. Our method efficiently segments different types of tissues in breast biopsy images while simultaneously predicting a discriminative map for identifying important areas in an image. Our network, Y-Net, extends and generalizes U-Net by adding a parallel branch for discriminative map generation and by supporting convolutional block modularity, which allows the user to adjust network efficiency without altering the network topology. Y-Net delivers state-of-the-art segmentation accuracy while learning 6.6x fewer parameters than its closest competitors. The addition of descriptive power from Y-Net's discriminative segmentation masks improve diagnostic classification accuracy by 7% over state-of-the-art methods for diagnostic classification. Source code is available at: https://sacmehta.github.io/YNet.
1 Introduction
Breast biopsy diagnosis depends on accurate interpretation, yet errors can harm patients and existing segmentation and classification methods provide complementary but incomplete information. Y-Net combines tissue-level segmentation with discriminative mapping to support simultaneous diagnosis.
- Diagnostic errors in breast biopsy interpretation can lead to incorrect treatment recommendations and significant patient harm.
- Whole-slide image analysis is difficult because of massive image size, motivating sliding-window approaches for medical-image classification and segmentation.
- Segmentation methods use tissue structure but face scarce expert labels and cannot weight tissue types according to diagnostic relevance.
- Y-Net simultaneously generates tissue-level segmentation masks and discriminative maps, combining segmentation and classification within one network.
- The study uses 428 breast-biopsy ROIs with diagnostic labels and tissue annotations from 58 ROIs to learn simultaneous segmentation and classification.
2 A System for Joint Segmentation and Classification
Y-Net extends U-Net with parallel segmentation and classification outputs, then combines them into a discriminative tissue-level mask for diagnosis. Its modular blocks and selection rule support efficient, adjustable processing of biopsy instances.
- Y-Net processes ROI instances and produces both instance-level tissue segmentation masks and diagnostic probability maps.
- Diagnostic probability maps are thresholded and combined with segmentation masks to form discriminative segmentation masks used for diagnosis.
- Y-Net Architecture: Y-Net generalizes U-Net by adding a classification branch distinct from the segmentation output.
- Y-Net Architecture: Modular encoding and decoding blocks allow different convolutional designs without changing topology, while width and depth multipliers vary network size.
- Discriminative Instance Selection: An instance is considered discriminative when the maximum softmax probability across diagnostic classes exceeds threshold τ.
- An MLP predicts cancer diagnosis from frequency and co-occurrence features extracted from the final discriminative mask.
3 Experiments
Experiments evaluate Y-Net’s modular design and compare it with state-of-the-art methods for tissue segmentation and diagnostic classification. The results show efficiency gains and improved classification from discriminative masks.
- Experimental Setup: The evaluation uses 428 ROIs with classification labels and 58 ROIs with tissue-level labels.
- Segmentation Results: The segmentation study trains on 30 ROIs and tests on 28, using 384 × 384 instances with overlap, augmentation, and multi-resolution extraction.
- Segmentation Results: Y-Net with ESP achieved similar segmentation performance to a state-of-the-art method while learning about 3× fewer parameters.
- Diagnostic Classification Results: Classification uses a 44-dimensional frequency-and-co-occurrence feature vector from discriminative masks to classify four diagnoses.
- Diagnostic Classification Results: About 9% higher classification accuracy resulted when discriminative masks replaced segmentation masks, with 62.5% accuracy approaching trained pathologists’ 70%.
4 Conclusion
On a breast biopsy dataset, Y-Net achieved strong segmentation and diagnostic classification performance. It matched state-of-the-art segmentation accuracy with fewer parameters and exceeded state-of-the-art diagnostic classification accuracy.
- Y-Net achieved good segmentation and diagnostic classification accuracy on a breast biopsy dataset.
- Y-Net matched state-of-the-art segmentation accuracy while learning fewer parameters.
- Discriminative segmentation-mask features enabled higher diagnostic classification accuracy than state-of-the-art methods.