Improved image classification explainability with high-accuracy heatmaps

an iScience paper

*Konpat Preechakul,  Sira Sriswasdi, Boonserm Kijsirikul,  Ekapol Chuangsuwanich*

PYLON’s heatmaps improve over traditional CAM heatmaps in both resolution and accuracy.

PYLON’s heatmaps improve over traditional CAM heatmaps in both resolution and accuracy.

Abstract

Deep learning models have become increasingly used for image-based classification. In critical applications such as medical imaging, it is important to convey the reasoning behind the models' decisions in human-understandable forms. In this work, we propose Pyramid Localization Network (PYLON), a deep learning model that delivers precise location explanation by increasing the resolution of heatmaps produced by class activation map (CAM). PYLON substantially improves the quality of CAM’s heatmaps in both general image and medical image domains and excels at pinpointing the locations of small objects. Most importantly, PYLON does not require expert annotation of the object location but instead can be trained using only image-level label. This capability is especially important for domain where expert annotation is often unavailable or costly to obtain. We also demonstrate an effective transfer learning approach for applying PYLON on small datasets and summarize technical guidelines that would facilitate wider adoption of the technique.

Paper

Untitled

Code

PYLON is ready to use with pretrained weights on Chest X-Ray 14 dataset.

PYLON is ready to use with pretrained weights on Chest X-Ray 14 dataset.

Highlights

A short tour

Class-activation maps (CAMs) techniques are widely used to better understand the decision made by conv nets. Showing as heatmaps, the explanation corresponds to the model’s region of interest.

One crucial limitation of CAM methods is low-resolution heatmaps leaving a lot to be desired if we are to use them in clinical settings.

In this work, we propose PYLON. PYLON is a deep model topping compatible with all modern conv net backbones such as ResNet, DenseNet, and EfficientNet.

Untitled