Connie Y Chang1, Colleen Buckless2, Kaitlyn J Yeh2, Martin Torriani2. 1. Division of Musculoskeletal Imaging and Intervention, Department of Radiology, Massachusetts General Hospital, Harvard Medical School, 55 Fruit Street, YAW 6, Boston, MA, 02114, USA. cychang@mgh.harvard.edu. 2. Division of Musculoskeletal Imaging and Intervention, Department of Radiology, Massachusetts General Hospital, Harvard Medical School, 55 Fruit Street, YAW 6, Boston, MA, 02114, USA.
Abstract
PURPOSE: To develop a deep convolutional neural network capable of detecting spinal sclerotic metastases on body CTs. MATERIALS AND METHODS: Our study was IRB-approved and HIPAA-compliant. Cases of confirmed sclerotic bone metastases in chest, abdomen, and pelvis CTs were identified. Images were manually segmented for 3 classes: background, normal bone, and sclerotic lesion(s). If multiple lesions were present on a slice, all lesions were segmented. A total of 600 images were obtained, with a 90/10 training/testing split. Images were stored as 128 × 128 pixel grayscale and the training dataset underwent a processing pipeline of histogram equalization and data augmentation. We trained our model from scratch on Keras/TensorFlow using an 80/20 training/validation split and a U-Net architecture (64 batch size, 100 epochs, dropout 0.25, initial learning rate 0.0001, sigmoid activation). We also tested our model's true negative and false positive rate with 1104 non-pathologic images. Global sensitivity measured model detection of any lesion on a single image, local sensitivity and positive predictive value (PPV) measured model detection of each lesion on a given image, and local specificity measured the false positive rate in non-pathologic bone. RESULTS: Dice scores were 0.83 for lesion, 0.96 for non-pathologic bone, and 0.99 for background. Global sensitivity was 95% (57/60), local sensitivity was 92% (89/97), local PPV was 97% (89/92), and local specificity was 87% (958/1104). CONCLUSION: A deep convolutional neural network has the potential to assist in detecting sclerotic spinal metastases.
PURPOSE: To develop a deep convolutional neural network capable of detecting spinal sclerotic metastases on body CTs. MATERIALS AND METHODS: Our study was IRB-approved and HIPAA-compliant. Cases of confirmed sclerotic bone metastases in chest, abdomen, and pelvis CTs were identified. Images were manually segmented for 3 classes: background, normal bone, and sclerotic lesion(s). If multiple lesions were present on a slice, all lesions were segmented. A total of 600 images were obtained, with a 90/10 training/testing split. Images were stored as 128 × 128 pixel grayscale and the training dataset underwent a processing pipeline of histogram equalization and data augmentation. We trained our model from scratch on Keras/TensorFlow using an 80/20 training/validation split and a U-Net architecture (64 batch size, 100 epochs, dropout 0.25, initial learning rate 0.0001, sigmoid activation). We also tested our model's true negative and false positive rate with 1104 non-pathologic images. Global sensitivity measured model detection of any lesion on a single image, local sensitivity and positive predictive value (PPV) measured model detection of each lesion on a given image, and local specificity measured the false positive rate in non-pathologic bone. RESULTS: Dice scores were 0.83 for lesion, 0.96 for non-pathologic bone, and 0.99 for background. Global sensitivity was 95% (57/60), local sensitivity was 92% (89/97), local PPV was 97% (89/92), and local specificity was 87% (958/1104). CONCLUSION: A deep convolutional neural network has the potential to assist in detecting sclerotic spinal metastases.
Authors: Shuling Li; Yi Peng; Eric D Weinhandl; Anne H Blaes; Karynsa Cetin; Victoria M Chia; Scott Stryker; Joseph J Pinzone; John F Acquavella; Thomas J Arneson Journal: Clin Epidemiol Date: 2012-04-10 Impact factor: 4.790
Authors: Olivier Q Groot; Michiel E R Bongers; Colleen G Buckless; Peter K Twining; Neal D Kapoor; Stein J Janssen; Joseph H Schwab; Martin Torriani; Miriam A Bredella Journal: J Surg Oncol Date: 2022-01-13 Impact factor: 2.885