Deep Learning Classifiers to Reduce False Positives in Osteolytic Lesion Segmentation Results from Low-dose CT Scans of Multiple Myeloma

Multiple myeloma (MM) is a hematological malignancy with a low survival rate if not detected in an early stage. A common symptom of MM is the development of osteolytic lesions, which appear as hypodense regions in bone tissue that are often visualised using low-dose CT imaging. Finding one lesion with a diamater of 5 mm or more is already enough to support the diagnosis of MM. However, evaluation of total-body CT (TBCT) scans is time consuming. Our group has developed an automated lesion segmentation algorithm to assist this process. Although providing accurate detection results, the algorithm results in excessive lesion-like false positive candidates. To address this problem
and further improve the segmentation performance, we deployed deep learning classifiers to reduce the false positive rate as a post-processing step. To train and evaluate the classifiers, a dataset was created, comprising of patches of lesions annotated by radiologists and images patches containing healthy bone tissue. The results showed that the best performing model, a fine-tuned ResNet50 model, achieved an F1 score of 0.83 on the test set. To test the performance of the model, expert radiologists labelled segmentation results as true or false positives for a hold-out test set. The model achieved an F1 score of 0.68 and a False Positive Rate (FPR) of 0.47 on the hold-out test set, reducing the number of false positives by 53%. By integrating our proposed model to the original segmentation pipeline, the number of reported false positives can be reduced, leading to a more reliable system.