Mingzhao Wang

Purpose: Breast cancer is the most common major public health problems of women in the world. Until now, analyzing mammogram images is still the main method used by doctors to diagnose and detect breast cancers. However, this process usually depends on the experience of radiologists and is always very time consuming. Patients and Methods: We propose to introduce deep learning technology into the process for the facilitation of computer-aided diagnosis (CAD), and address the challenges of class imbalance, enhance the detection of small masses and multiple targets, and reduce false positives and negatives in mammogram analysis. Therefore, we adopted and enhanced RetinaNet to detect masses in mammogram images. Specifically, we introduced a novel modification to the network structure, where the feature map M5 is processed by the ReLU function prior to the original convolution kernel. This strategic adjustment was designed to prevent the loss of resolution for small mass features. Additionally, we introduced transfer learning techniques into training process through leveraging pre-trained weights from other RetinaNet applications, and fine-tuned our improved model using the INbreast dataset. Results: The aforementioned innovations facilitate superior performance of the enhanced RetiaNet model on the public dataset INbreast, as evidenced by a mAP (mean average precision) of 1.0000 and TPR (true positive rate) of 1.00 at 0.00 FPPI (false positive per image) on the INbreast dataset. Conclusion: The experimental results demonstrate that our enhanced RetinaNet model defeats the existing models by having more generalization performance than other published studies, and it can also be applied to other types of patients to assist doctors in making a proper diagnosis.

Head and neck cancer is one of the most common malignancies in the world. We propose SE-nnU-Net by adapting SE (squeeze and excitation) normalization into nnU-Net, so as to segment head and neck tumors in PET/CT images by combining advantages of SE capturing features of interest regions and nnU-Net configuring itself for a specific task. The basic module referred to convolution-ReLU-SE is designed for SE-nnU-Net. In the encoder it is combined with residual structure while in the decoder without residual structure. The loss function combines Dice loss and Focal loss. The specific data preprocessing and augmentation techniques are developed, and specific network architecture is designed. Moreover, the deep supervised mechanism is introduced to calculate the loss function using the last four layers of the decoder of SE-nnU-Net. This SE-nnU-net is applied to HECKTOR 2020 and HECKTOR 2021 challenges, respectively, using different experimental design. The experimental results show that SE-nnU-Net for HECKTOR 2020 obtained 0.745, 0.821, and 0.725 in terms of Dice, Precision, and Recall, respectively, while the SE-nnU-Net for HECKTOR 2021 obtains 0.778 and 3.088 in terms of Dice and median HD95, respectively. This SE-nnU-Net for segmenting head and neck tumors can provide auxiliary opinions for doctors' diagnoses.