Towards Interpretable Semantic Segmentation via Gradient-Weighted Class Activation Mapping (Student Abstract)

Vinogradova, Kira; Dibrov, Alexandr; Myers, Gene

doi:10.1609/aaai.v34i10.7244

Cited by 84 publications

(45 citation statements)

References 6 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…We generally find the convolutional layers near the end of U-Net's contracting path captures more comprehensible features while the convolutional layers in the expansive path successively combine features from the contracting path and produce heatmaps that look more and more similar to the logits of the selected class. This Grad-CAM finding is supported by similar results of U-Net semantic segmentation on Cityscapes datasets [57]. In Fig.…”

Section: ) Resnet-18supporting

confidence: 78%

Automated Segmentation of Amyloid-β Stained Whole Slide Images of Brain Tissue

Lai

Guo

et al. 2020

Preprint

View full text Add to dashboard Cite

Neurodegenerative disease pathologies have been reported in both grey matter (GM) and white matter (WM) with different density distributions, an automated separation task of GM/WM would be extremely advantageous for aid in neuropathologic deep phenotyping. Standard segmentation methods typically involve manual annotations, where a trained researcher traces the delineation of GM/WM in ultra-high-resolution Whole Slide Images (WSIs). This method can be time-consuming and subjective, preventing the analysis of large amounts of WSIs in a scalable way. In this paper, we propose an automated segmentation pipeline combining a Convolutional Neural Network (CNN) module for segmenting GM/WM regions and a post-processing module to remove artifacts/residues of tissues as well as generate XML annotations that can be visualized via Aperio ImageScope. First, we investigate two baseline models for medical image segmentation: FCN, and U-Net. Then we propose a new patch-based approach, ResNet-Patch, to classify the GM/WM/background regions. In addition, we integrate a Neural Conditional Random Field (NCRF) module, ResNet-NCRF, to model and incorporate the spatial correlations among neighboring patches. Although their mechanisms are greatly different, both U-Net and ResNet-Patch/ResNet-NCRF achieve Intersection over Union (IoU) of more than 90% in GM and more than 80% in WM, while ResNet-Patch achieves 1% superior to U-Net with lower variance among various WSIs. ResNet-NCRF further improves the IoU by 3% for WM compared to ResNet-Patch before post-processing. We also apply gradient-weighted class activation mapping (Grad-CAM) to interpret the segmentation masks and provide clinical explanations and insights.

show abstract

Section: ) Resnet-18supporting

confidence: 78%

Automated Segmentation of Amyloid-β Stained Whole Slide Images of Brain Tissue

Lai

Guo

et al. 2020

Preprint

View full text Add to dashboard Cite

show abstract

“…For data interpretability, a semantic segmentation extension of the gradient class activation mapping (GradCAM++) strategy was applied [43], which aims to visually test the deep learning model's ability to learn relevant features separating between nerve and background. Specifically, a Grad-CAM++ extension of the seminal work in [44] for semantic segmentation was proposed to capture the entire object completeness. Then, an explanation map-based quantitative assessment was carried out for relevance analysis.…”

Section: Introductionmentioning

confidence: 99%

Random Fourier Features-Based Deep Learning Improvement with Class Activation Interpretability for Nerve Structure Segmentation

Jimenez-Castaño

Álvarez-Meza

Aguirre-Ospina

et al. 2021

Sensors

View full text Add to dashboard Cite

Peripheral nerve blocking (PNB) is a standard procedure to support regional anesthesia. Still, correct localization of the nerve’s structure is needed to avoid adverse effects; thereby, ultrasound images are used as an aid approach. In addition, image-based automatic nerve segmentation from deep learning methods has been proposed to mitigate attenuation and speckle noise ultrasonography issues. Notwithstanding, complex architectures highlight the region of interest lacking suitable data interpretability concerning the learned features from raw instances. Here, a kernel-based deep learning enhancement is introduced for nerve structure segmentation. In a nutshell, a random Fourier features-based approach was utilized to complement three well-known semantic segmentation architectures, e.g., fully convolutional network, U-net, and ResUnet. Moreover, two ultrasound image datasets for PNB were tested. Obtained results show that our kernel-based approach provides a better generalization capability from image segmentation-based assessments on different nerve structures. Further, for data interpretability, a semantic segmentation extension of the GradCam++ for class-activation mapping was used to reveal relevant learned features separating between nerve and background. Thus, our proposal favors both straightforward (shallow) and complex architectures (deeper neural networks).

show abstract

“…There are only a few previous works that focus on segmentation interpretability [6,7]. These works solve the problem in a lower resolution setup, and then up-sample to get pixel level explanation.…”

Section: Introductionmentioning

confidence: 99%

“…These works solve the problem in a lower resolution setup, and then up-sample to get pixel level explanation. They also only return explanations with respect to a single pixel [6] or a part of the image [7]. Our method on the other hand, seamlessly returns the relative importance of each pixel, with respect to the entire input.…”

Section: Introductionmentioning

confidence: 99%

U-Noise: Learnable Noise Masks for Interpretable Image Segmentation

Koker¹,

Mireshghallah

Titcombe³

et al. 2021

2021 IEEE International Conference on Image Processing (ICIP)

View full text Add to dashboard Cite

Deep Neural Networks (DNNs) are widely used for decision making in a myriad of critical applications, ranging from medical to societal and even judicial. Given the importance of these decisions, it is crucial for us to be able to interpret these models. We introduce a new method for interpreting image segmentation models by learning regions of images in which noise can be applied without hindering downstream model performance. We apply this method to segmentation of the pancreas in CT scans, and qualitatively compare the quality of the method to existing explainability techniques, such as Grad-CAM and occlusion sensitivity. Additionally we show that, unlike other methods, our interpretability model can be quantitatively evaluated based on the downstream performance over obscured images.

show abstract

Towards Interpretable Semantic Segmentation via Gradient-Weighted Class Activation Mapping (Student Abstract)

Cited by 84 publications

References 6 publications

Automated Segmentation of Amyloid-β Stained Whole Slide Images of Brain Tissue

Automated Segmentation of Amyloid-β Stained Whole Slide Images of Brain Tissue

Random Fourier Features-Based Deep Learning Improvement with Class Activation Interpretability for Nerve Structure Segmentation

U-Noise: Learnable Noise Masks for Interpretable Image Segmentation

Contact Info

Product

Resources

About