Editorial for “Automated Segmentation of Brain Metastases on T1‐Weighted MRI Using Convolutional Neural Network: Impact of Using Volume Aware Loss and Sampling Strategy”
Editorial for “Automated Segmentation of Brain Metastases on T1-Weighted MRI Using Convolutional Neural Network: Impact of Using Volume Aware Loss and Sampling Strategy” Using deep learning convolutional neural networks for brain metastasis segmentation in MRI has been a hot topic because a reliable and automated deep learning algorithm with high sensitivity and accuracy can greatly benefit treatment planning for stereotactic radiosurgery. During the treatment planning, brain metastases are currently manually detected in MRI by experienced radiologists and segmented by radiation oncologists, which are time-consuming and labor-intensive. A standardized algorithm helping with detection and segmentation can not only improve treatment planning efficiency and avoid miss-detection, but also reduce inter-practitioner variability for brain metastasis segmentation. To achieve the goal of developing a clinically helpful segmentation algorithm, multiple deep learning-based studies with MRIs have been performed. In most of the studies, researchers used two approaches to brain metastasis segmentation. The first approach is direct segmentation of brain metastases from MRI images or patches. The deep neural networks developed in this approach include DeepMedic, EnDeepMedic, and a modified GoogLeNet. The second approach is to first detect brain metastases on images and then segment the lesions based on the detection. The deep learning frameworks developed in this approach include a UNet cascade, in which the first U-Net is used to detect brain metastases and the second is used to segment the detected lesions. In another study, investigators used single-shot detector for brain metastasis detection and then used a U-Net for segmentation of the metastases. Among the studies, a common challenge in further improving the detection and segmentation performance is the balance between the detection sensitivity and false positives, especially for small metastases with diameters less than 3 mm. The detection sensitivity for small metastases is limited, and the incidence of false positives per patient will increase if the sensitivity improves. In this study, the authors aimed to improve the detection sensitivity for small lesions while keeping the number of false positives per patient relatively low. By using a U-Netlike network, they directly segmented brain metastases on postcontrast T1-weighted MRIs. Besides the residual connections in the network, the authors proposed to incorporate novel loss functions such as volume aware (VA) loss and boundary loss, and different sampling strategies to train the constructed convolutional neural network. VA loss is a modified Dice similarity coefficient loss. By incorporating a weighting factor that is inversely proportional to the lesion volume, this new Dice coefficient has greater contribution from smaller lesions, alleviating the imbalance between small and large lesions from the training batches. Additionally, the authors incorporated boundary loss with the Dice coefficient for training. The boundary loss was used to penalize the lesion predictions that were away from the ground truth contours, suppressing the false positive predictions. Also, to improve the network awareness of small lesions, the authors used VA sampling during network training. VA sampling uses a probability map that first assigns the sampling probability equally among all lesions. Then, for each pixel within each lesion, the sampling probability is inverse proportional to the lesion volume. Therefore, pixels within small lesions will have greater probability of being sampled for network training. Furthermore, by adjusting the sampling probability for the background pixels, which typically have a much greater number than lesion pixels, the VA sampling strategy can control the number of background pixels sampled during network training. With the combination of VA loss and VA sampling, the authors achieved a peak overall sensitivity of 91%, with 0.66 false positives per patient. For the lesions smaller than 6 mm but larger than 2.5 mm in diameter, their network achieved a sensitivity of 80%, with 0.55 false positives per patient.
Paper
Full text
Editorial for “Automated Segmentation of Brain Metastases on T1‐Weighted MRI Using Convolutional Neural Network: Impact of Using Volume Aware Loss and Sampling Strategy”
Semantic Scholar · Medicine · 2022
Abstract
Editorial for “Automated Segmentation of Brain Metastases on T1-Weighted MRI Using Convolutional Neural Network: Impact of Using Volume Aware Loss and Sampling Strategy” Using deep learning convolutional neural networks for brain metastasis segmentation in MRI has been a hot topic because a reliable and automated deep learning algorithm with high sensitivity and accuracy can greatly benefit treatment planning for stereotactic radiosurgery. During the treatment planning, brain metastases are currently manually detected in MRI by experienced radiologists and segmented by radiation oncologists, which are time-consuming and labor-intensive. A standardized algorithm helping with detection and segmentation can not only improve treatment planning efficiency and avoid miss-detection, but also reduce inter-practitioner variability for brain metastasis segmentation. To achieve the goal of developing a clinically helpful segmentation algorithm, multiple deep learning-based studies with MRIs have been performed. In most of the studies, researchers used two approaches to brain metastasis segmentation. The first approach is direct segmentation of brain metastases from MRI images or patches. The deep neural networks developed in this approach include DeepMedic, EnDeepMedic, and a modified GoogLeNet. The second approach is to first detect brain metastases on images and then segment the lesions based on the detection. The deep learning frameworks developed in this approach include a UNet cascade, in which the first U-Net is used to detect brain metastases and the second is used to segment the detected lesions. In another study, investigators used single-shot detector for brain metastasis detection and then used a U-Net for segmentation of the metastases. Among the studies, a common challenge in further improving the detection and segmentation performance is the balance between the detection sensitivity and false positives, especially for small metastases with diameters less than 3 mm. The detection sensitivity for small metastases is limited, and the incidence of false positives per patient will increase if the sensitivity improves. In this study, the authors aimed to improve the detection sensitivity for small lesions while keeping the number of false positives per patient relatively low. By using a U-Netlike network, they directly segmented brain metastases on postcontrast T1-weighted MRIs. Besides the residual connections in the network, the authors proposed to incorporate novel loss functions such as volume aware (VA) loss and boundary loss, and different sampling strategies to train the constructed convolutional neural network. VA loss is a modified Dice similarity coefficient loss. By incorporating a weighting factor that is inversely proportional to the lesion volume, this new Dice coefficient has greater contribution from smaller lesions, alleviating the imbalance between small and large lesions from the training batches. Additionally, the authors incorporated boundary loss with the Dice coefficient for training. The boundary loss was used to penalize the lesion predictions that were away from the ground truth contours, suppressing the false positive predictions. Also, to improve the network awareness of small lesions, the authors used VA sampling during network training. VA sampling uses a probability map that first assigns the sampling probability equally among all lesions. Then, for each pixel within each lesion, the sampling probability is inverse proportional to the lesion volume. Therefore, pixels within small lesions will have greater probability of being sampled for network training. Furthermore, by adjusting the sampling probability for the background pixels, which typically have a much greater number than lesion pixels, the VA sampling strategy can control the number of background pixels sampled during network training. With the combination of VA loss and VA sampling, the authors achieved a peak overall sensitivity of 91%, with 0.66 false positives per patient. For the lesions smaller than 6 mm but larger than 2.5 mm in diameter, their network achieved a sensitivity of 80%, with 0.55 false positives per patient.