Development of the algorithm for differentiating bone metastases and trauma of the ribs in bone scintigraphy and demonstration of visual evidence of the algorithm -- Using only anterior bone scan view of thorax

TL;DR

This study developed an AI algorithm using only anterior thorax bone scans to differentiate rib bone metastases from trauma, visualizing the diagnostic focus with Grad-CAM, achieving high accuracy and sensitivity.

Contribution

The paper introduces a novel AI model that classifies rib lesions as metastasis or trauma using only anterior bone scans and visualizes diagnostic areas with Grad-CAM.

Findings

Achieved 90% sensitivity in detecting metastases.

Reached 86.5% overall accuracy.

Successfully visualized AI focus areas with Grad-CAM.

Abstract

Background: Although there are many studies on the application of artificial intelligence (AI) models to medical imaging, there is no report of an AI model that determines the accumulation of ribs in bone metastases and trauma only using the anterior image of thorax of bone scintigraphy. In recent years, a method for visualizing diagnostic grounds called Gradient-weighted Class Activation Mapping (Grad-CAM) has been proposed in the area of diagnostic images using Deep Convolutional Neural Network (DCNN). As far as we have investigated, there are no reports of visualization of the diagnostic basis in bone scintigraphy. Our aim is to visualize the area of interest of DCNN, in addition to developing an algorithm to classify and diagnose whether RI accumulation on the ribs is bone metastasis or trauma using only anterior bone scan view of thorax. Material and Methods: For this retrospective study, we used 838 patients who underwent bone scintigraphy to search for bone metastases at our institution. A frontal chest image of bone scintigraphy was used to create the algorithm. We used 437 cases with bone metastases on the ribs and 401 cases with abnormal RI accumulation due to trauma. Result: AI model was able to detect bone metastasis lesion with a sensitivity of 90.00% and accuracy of 86.5%. And it was possible to visualize the part that the AI model focused on with Grad-CAM.