Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Deep learning for automated detection of breast cancer in deep ultraviolet fluorescence images with diffusion probabilistic model

Sepehr Salem Ghahfarokhi, Tyrell To, Julie Jorns, Tina Yen, Bing Yu, Dong Hye Ye

TL;DR

This work tackles data scarcity in intraoperative breast cancer margin assessment using deep ultraviolet fluorescence images by employing a diffusion probabilistic model (DPM) to synthesize labeled patches for augmentation. A patch-based pipeline extracts ResNet50 features, classifies with XGBoost, and fuses patch decisions with Grad-CAM++–derived regional maps to yield whole-surface predictions. DPM augmentation substantially improves performance, achieving 97% accuracy with 97% sensitivity and 93% specificity, outperforming affine augmentation and ProGAN baselines. The study demonstrates that diffusion-based data augmentation can enhance robustness and generalization in breast cancer detection from scarce medical imaging data, with practical implications for reducing margin misclassification during breast-conserving surgery.

Abstract

Data limitation is a significant challenge in applying deep learning to medical images. Recently, the diffusion probabilistic model (DPM) has shown the potential to generate high-quality images by converting Gaussian random noise into realistic images. In this paper, we apply the DPM to augment the deep ultraviolet fluorescence (DUV) image dataset with an aim to improve breast cancer classification for intraoperative margin assessment. For classification, we divide the whole surface DUV image into small patches and extract convolutional features for each patch by utilizing the pre-trained ResNet. Then, we feed them into an XGBoost classifier for patch-level decisions and then fuse them with a regional importance map computed by Grad-CAM++ for whole surface-level prediction. Our experimental results show that augmenting the training dataset with the DPM significantly improves breast cancer detection performance in DUV images, increasing accuracy from 93% to 97%, compared to using Affine transformations and ProGAN.

Deep learning for automated detection of breast cancer in deep ultraviolet fluorescence images with diffusion probabilistic model

TL;DR

This work tackles data scarcity in intraoperative breast cancer margin assessment using deep ultraviolet fluorescence images by employing a diffusion probabilistic model (DPM) to synthesize labeled patches for augmentation. A patch-based pipeline extracts ResNet50 features, classifies with XGBoost, and fuses patch decisions with Grad-CAM++–derived regional maps to yield whole-surface predictions. DPM augmentation substantially improves performance, achieving 97% accuracy with 97% sensitivity and 93% specificity, outperforming affine augmentation and ProGAN baselines. The study demonstrates that diffusion-based data augmentation can enhance robustness and generalization in breast cancer detection from scarce medical imaging data, with practical implications for reducing margin misclassification during breast-conserving surgery.

Abstract

Data limitation is a significant challenge in applying deep learning to medical images. Recently, the diffusion probabilistic model (DPM) has shown the potential to generate high-quality images by converting Gaussian random noise into realistic images. In this paper, we apply the DPM to augment the deep ultraviolet fluorescence (DUV) image dataset with an aim to improve breast cancer classification for intraoperative margin assessment. For classification, we divide the whole surface DUV image into small patches and extract convolutional features for each patch by utilizing the pre-trained ResNet. Then, we feed them into an XGBoost classifier for patch-level decisions and then fuse them with a regional importance map computed by Grad-CAM++ for whole surface-level prediction. Our experimental results show that augmenting the training dataset with the DPM significantly improves breast cancer detection performance in DUV images, increasing accuracy from 93% to 97%, compared to using Affine transformations and ProGAN.
Paper Structure (9 sections, 7 equations, 2 figures, 1 table)

This paper contains 9 sections, 7 equations, 2 figures, 1 table.

Table of Contents

  1. Introduction
  2. METHOD
  3. Diffusion Probabilistic Model
  4. Deep learning classification of DUV images
  5. Experiment
  6. Dataset
  7. Visual Inspection
  8. Classification Performance
  9. CONCLUSION

Figures (2)

  • Figure 1: Overview of the proposed method: The proposed method starts by extracting patches from Whole Surface Images (WSI-DUV). A two-step diffusion process, involving noise addition and removal with probabilistic models, is applied to generate patch images. Utilizing a generated dataset alongside an existing training dataset, deep convolutional features are extracted using a pre-trained ResNet50 network. Patch-level classification is then performed using XGBoost, and a regional importance map is computed with Grad-CAM++ on a pre-trained DenseNet169 model for DUV-WSI. The final prediction at the WSI level is achieved through a decision fusion approach, combining patch-level results with the regional importance map.
  • Figure 2: Comparison of synthesized patch images using ProGAN and DPM: The images generated by DPM closely mimic real biological features, showcasing characteristics like enlarged cells, dense cellular structures, infiltration, and varied nuclear traits in both malignant and benign types, in contrast to ProGAN (a: malignant, b: benign).it's apparent that DPM produces high-quality, sharp data samples with intricate features, while ProGAN generates images with noticeable blurring artifacts. Additionally, our method stands out due to its capability to capture a diverse range of image variations.