TY - CHAP
T1 - Deep Learning Classifiers to Reduce False Positives in Osteolytic Lesion Segmentation Results from Low-Dose CT Scans of Multiple Myeloma
AU - van Leeuwen, Martijn
AU - Jadikar, Munirdin
AU - van Oudheusden, Thijs
AU - Oei, Sebastian
AU - Steunenberg, Bastiaan
AU - Kint, Rik
AU - Ranschaert, Erik
AU - Bosma, Gerlof
AU - Saygili, Görkem
AU - Ong, Sharon
PY - 2024/11/2
Y1 - 2024/11/2
N2 - Multiple myeloma (MM) is a hematological malignancy with a low survival rate if not diagnosed in an early stage. A common symptom of MM is the development of osteolytic lesions, which are often visualized using low-dose CT imaging. Finding one lesion with a diameter of 5 mm or more is enough to support the diagnosis of MM. However, the evaluation of whole-body CT scans (WBCT) is time-consuming and susceptible to errors. We developed an automated lesion segmentation algorithm to assist in the detection and segmentation of such lesions. Although providing adequate detection results, this model returned false positive predictions which hampers the application of the method in the clinic. This work utilized various deep learning classifiers to categorize the segmentation outputs into true positives or false positives. A dataset comprising patches containing osteolytic lesions, and image patches containing regular bone tissue was constructed. After training on this dataset, 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 further test generalizability, radiologists labeled a collection of segmentation results as true or false positives, creating a separate holdout test set. The model achieved an F1 score of 0.68 a False Positive Rate (FPR) of 0.47 and a False Negative Rate (FNR) of 0.24 on the holdout test set. Our results show the number of false positive segmentations can be reduced by 53% while still detecting 76% of the lesions. Implementing our proposed model into the segmentation pipeline can reduce the number of false positive segmentation, leading to a more refined and reliable system. However, further research is required to optimize the model performance.
AB - Multiple myeloma (MM) is a hematological malignancy with a low survival rate if not diagnosed in an early stage. A common symptom of MM is the development of osteolytic lesions, which are often visualized using low-dose CT imaging. Finding one lesion with a diameter of 5 mm or more is enough to support the diagnosis of MM. However, the evaluation of whole-body CT scans (WBCT) is time-consuming and susceptible to errors. We developed an automated lesion segmentation algorithm to assist in the detection and segmentation of such lesions. Although providing adequate detection results, this model returned false positive predictions which hampers the application of the method in the clinic. This work utilized various deep learning classifiers to categorize the segmentation outputs into true positives or false positives. A dataset comprising patches containing osteolytic lesions, and image patches containing regular bone tissue was constructed. After training on this dataset, 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 further test generalizability, radiologists labeled a collection of segmentation results as true or false positives, creating a separate holdout test set. The model achieved an F1 score of 0.68 a False Positive Rate (FPR) of 0.47 and a False Negative Rate (FNR) of 0.24 on the holdout test set. Our results show the number of false positive segmentations can be reduced by 53% while still detecting 76% of the lesions. Implementing our proposed model into the segmentation pipeline can reduce the number of false positive segmentation, leading to a more refined and reliable system. However, further research is required to optimize the model performance.
U2 - 10.1007/978-3-031-74650-5_9
DO - 10.1007/978-3-031-74650-5_9
M3 - Chapter
SN - 978-3-031-74649-9
VL - 2187
T3 - Communications in Computer and Information Science
SP - 155
EP - 173
BT - Artificial Intelligence and Machine Learning. BNAIC/Benelearn 2023
A2 - Oliehoek, Frans
A2 - Kok, Manon
A2 - Verwer, Sicco
PB - Springer Cham
ER -