Source-linked AI summary

A Multi-Scale CNN and Curriculum Learning Strategy for Mammogram Classification

William Lotter, Greg Sorensen, David Cox

arXiv:1707.06978v1cs.CV

TL;DR

Full-image mammogram classification is difficult because small, subtle lesions create a “needle in a haystack” problem, while existing approaches often target pipeline subtasks. The paper uses lesion-specific, multi-scale patch classifiers to initialize a scanning-window CNN trained end-to-end, achieving 0.92 AUROC on DDSM. The results support this architecture and curriculum strategy as a promising approach for mammogram classification, while future work seeks to reduce reliance on hand-drawn segmentation masks.

  • Problem

    Full-image mammogram cancer-status classification is challenging because small, subtle lesions must be detected in high-resolution images, and many prior approaches address only individual pipeline subtasks.

  • Method

    The method pretrains separate multi-scale lesion patch classifiers, uses their features in a scanning-window image model, and fine-tunes the system end-to-end on binary cancer labels.

  • Results

    0.92 AUROC was achieved on the DDSM dataset against final pathology outcomes.

  • Takeaways & Limitations

    The multi-scale CNN and lesion-specific curriculum strategy provide a promising approach for full-image mammogram classification and outperform standard out-of-the-box CNN models.

Abstract

from arXiv · show

Screening mammography is an important front-line tool for the early detection of breast cancer, and some 39 million exams are conducted each year in the United States alone. Here, we describe a multi-scale convolutional neural network (CNN) trained with a curriculum learning strategy that achieves high levels of accuracy in classifying mammograms. Specifically, we first train CNN-based patch classifiers on segmentation masks of lesions in mammograms, and then use the learned features to initialize a scanning-based model that renders a decision on the whole image, trained end-to-end on outcome data. We demonstrate that our approach effectively handles the "needle in a haystack" nature of full-image mammogram classification, achieving 0.92 AUROC on the DDSM dataset.

1 Introduction

Screening mammography supports early breast-cancer detection but remains difficult to interpret reliably because lesions are small and subtle, readings are tedious, and false positives impose costs. The paper addresses whole-image cancer-status classification, a challenging “needle in a haystack” problem, with a two-stage deep-learning approach achieving 0.92 AUROC on DDSM.

  • Screening mammography is important for early breast-cancer detection, but interpretation is tedious, error-prone, and associated with false positives, cost, and patient stress.
  • Many prior deep-learning approaches address individual mammography subtasks rather than binary cancer-status classification from an entire mammogram.
  • The proposed curriculum strategy trains multi-scale patch classifiers before using their features to initialize an end-to-end image-level model.
  • 0.92 AUROC was achieved on DDSM against final pathology outcomes.

2 Multi-Scale CNN with Curriculum Learning Strategy

The method uses lesion-specific, multi-scale patch learning to initialize a scanning-window model for full-image classification. It combines patch features across scales through pooling and concatenation, then fine-tunes the complete system end-to-end.

  • Motivation: Lesions occupy small image regions, with median masks covering about 0.5% of calcification images and 1.2% of mass images.The high-resolution DDSM images require detecting fine detail in small regions.
  • Patch-level curriculum: The first stage trains separate Wide ResNet patch classifiers for calcifications and masses using segmentation-derived labels.Training proceeds from abnormality detection to pathology-determined malignancy.
  • Image-level training: The second stage tiles each mammogram with minimally overlapping patches and extracts features using the trained patch models.
  • Image-level training: Global average pooling aggregates patch features at each scale, followed by concatenation and a softmax classifier.
  • Image-level training: The image-level model is fine-tuned end-to-end using binary image-level labels.

3 Experiments

Experiments evaluate the model on patient-separated DDSM data using breast-level predictions and compare it with InceptionV3 and ablated training choices. The proposed model reaches 0.92 ± 0.02 AUC, while removing size augmentation or lesion pre-training reduces performance.

  • Dataset and evaluation: DDSM contains 10,480 images from 2,620 cases, split into 87% training, 5% validation, and 8% testing by patient.Each case includes CC and MLO views for both breasts.
  • Training setup: The image-level model uses initialized patch weights, multi-scale 256x256 patches, overlap normalization, and malignant-lesion labels aggregated across breast views.
  • Baseline: InceptionV3 cannot directly use random input-size augmentation because it assumes a fixed input size.
  • Ablations: 0.89 ± 0.02 AUC resulted without size-augmentation training, compared with 0.92 ± 0.02 AUC for the full model.
  • Baselines and ablations: 0.65 ± 0.04 AUC resulted without DDSM lesion pre-training, while InceptionV3 achieved 0.77 ± 0.03 with and 0.59 ± 0.04 without ImageNet pre-training.

4 Conclusions

The paper presents a multi-scale CNN with lesion-specific curriculum learning for full-image mammogram classification. It reports promising performance relative to standard CNNs while identifying architecture and training strategy as important components and proposing less reliance on hand-drawn masks as future work.

  • The proposed multi-scale scanning-window CNN with lesion-specific curriculum learning achieves promising full-image mammogram classification results.
  • The approach performs significantly better than standard out-of-the-box CNN models.
  • Experiments indicate that both architecture choice and training scheme are important for performance.
  • Future work aims to learn interest points more unsupervisedly to reduce reliance on hand-drawn segmentation masks.
Loading 1707.06978v1…