Deep learning based identification of bone scintigraphies containing metastatic bone disease foci

Abstract

Purpose: Metastatic bone disease (MBD) is the most common form of metastases, most frequently deriving from prostate cancer. MBD is screened with bone scintigraphy (BS), which have high sensitivity but low specificity for the diagnosis of MBD, often requiring further investigations. Deep learning (DL) - a machine learning technique designed to mimic human neuronal interactions- has shown promise in the field of medical imaging analysis for different purposes, including segmentation and classification of lesions. In this study, we aim to develop a DL algorithm that can classify areas of increased uptake on bone scintigraphy scans.

Methods: We collected 2365 BS from three European medical centres. The model was trained and validated on 1203 and 164 BS scans respectively. Furthermore we evaluated its performance on an external testing set composed of 998 BS scans. We further aimed to enhance the explainability of our developed algorithm, using activation maps. We compared the performance of our algorithm to that of 6 nuclear medicine physicians.

Results: The developed DL based algorithm is able to detect MBD on BSs, with high specificity and sensitivity (0.80 and 0.82 respectively on the external test set), in a shorter time compared to the nuclear medicine physicians (2.5 min for AI and 30 min for nuclear medicine physicians to classify 134 BSs). Further prospective validation is required before the algorithm can be used in the clinic.

Keywords: Activation maps; Bone scintigraphy; Deep learning; Metastatic bone disease.

Conflict of interest statement

Akshayaa Vaidyanathan, Flore Belmans, Fabio Bottari are salaried employees of Radiomics.

The rest of co-authors declare no competing interest.

© 2023. The Author(s).

Figures

Fig. 1

Example of pre-processed BS scans used as input for model training

Fig. 2

The architecture used in the study. Pre-processed BS scans resized to 512 * 512 dimensions were provided as input to the network. The network outputs a probability score for presence and absence of metastasis on BS images. X = block repetitions, Conv = Convolution kernel, ReLU = rectified linear unit, 3 × 3 = the size of the 2D CNN kernels

Fig. 3

Screenshot of the application feedback window used in the in silico trial

Fig. 4

ROC curve for the classification DL model (left) and Confusion matrix (right)

Fig. 5

BS images which are correctly classified along with their corresponding activation maps extracted using the GRAD-CAM method. Left) original BD scan, Right) Grad-CAM activation maps obtained from the DL model. Scan correctly classified with a probability of 0.78 (top) and 0.99 (bottom)

Fig. 6

BS images which are wrongly classified along with their corresponding activation maps extracted using the GRAD-CAM method. Left) original BD scan, Right) Grad-CAM activation maps obtained from the DL model. Scan incorrectly classified with a probability of 0.79 (top) and 0.63 (bottom)

Fig. 7

Violin plots showing the distributions of AUC scores for DL based and manual (across physicians) metastases detection on BS (left); boxplots of the log of the time needed by DL algorithm and nuclear medicine physicians (right)