Explainable AI for CNN-based Prostate Tumor Segmentation in Multi-parametric MRI Correlated to Whole Mount Histopathology

Automatic prostate tumor segmentation is often unable to identify the lesion even if in multi-parametric MRI data is used as input, and the segmentation output is difficult to verify due to the lack of clinically established ground truth images. In this work we use an explainable deep learning model to interpret the predictions of a convolutional neural network (CNN) for prostate tumor segmentation. The CNN uses a U-Net architecture which was trained on multi-parametric MRI data from 122 patients to automatically segment the prostate gland and prostate tumor lesions. In addition, co-registered ground truth data from whole mount histopathology images were available in 15 patients that were used as a test set during CNN testing. To be able to interpret the segmentation results of the CNN, heat maps were generated using the Gradient Weighted Class Activation Map (Grad-CAM) method. With the CNN a mean Dice Sorensen Coefficient for the prostate gland and the tumor lesions of 0.62 and 0.31 with the radiologist drawn ground truth and 0.32 with wholemount histology ground truth for tumor lesions could be achieved. Dice Sorensen Coefficient between CNN predictions and manual segmentations from MRI and histology data were not significantly different. In the prostate the Grad-CAM heat maps could differentiate between tumor and healthy prostate tissue, which indicates that the image information in the tumor was essential for the CNN segmentation.

Weakly Supervised Learning for Object Localization Based on an Attention Mechanism

Recently, deep learning has been successfully applied to object detection and localization tasks in images. When setting up deep learning frameworks for supervised training with large datasets, strongly labeling the objects facilitates good performance; however, the complexity of the image scene and large size of the dataset make this a laborious task. Hence, it is of paramount importance that the expensive work associated with the tasks involving strong labeling, such as bounding box annotation, is reduced. In this paper, we propose a method to perform object localization tasks without bounding box annotation in the training process by means of employing a two-path activation-map-based classifier framework. In particular, we develop an activation-map-based framework to judicially control the attention map in the perception branch by adding a two-feature extractor so that better attention weights can be distributed to induce improved performance. The experimental results indicate that our method surpasses the performance of the existing deep learning models based on weakly supervised object localization. The experimental results show that the proposed method achieves the best performance, with 75.21% Top-1 classification accuracy and 55.15% Top-1 localization accuracy on the CUB-200-2011 dataset.

Local impedance characteristics and advanced mapping capabilities to better understand pulmonary veins reconnections during repeat AF ablation procedures: insight from the CHARISMA registry

Background Detailed characterization of pulmonary veins (PV) reconnection during repeat AF ablation through high-density mapping (HDM) and local impedance (LI) algorithm is still lacking. Purpose We aimed to characterize PV gaps and underlying electrical activity during and after ablation of PVs in AF patients (pts). Methods Consecutive patients (pts) undergoing redo AF ablation from the CHARISMA registry with complete characterization of PV gaps (PVG) at 8 Italian centers were included. Rhythmia mapping system was used to map the left atrium and PVs before and after ablation. LI characteristics were collected through a RF ablation catheter equipped with a dedicated LI algorithm (DirectSense). A novel map analysis tool (Lumipoint) that automatically identifies split potentials and continuous activation was used sequentially on each PV component, in order to better assess PVG. Each PVG was characterized in terms of LI and its variations during the procedure. Ablation endpoint was PVI as assessed by entrance and exit block. Results Fifty PVGs were automatically identified through the Lumipoint tool in 23 cases, mostly at anterior sites (21, 42%), followed by posterior (15, 30%) and carina (10, 20%) sites. One PVG was identified in 7 (28%) pts, 2 gaps in 10 (43.5%) pts and >2 gaps in 6 (26.1%) pts. The mean LI at PVG sites was 111.3±12Ω prior to ablation: it was significantly higher than LI at scar tissue closer to PVG (99.3±8Ω, p<0.0001) but was significantly lower than LI at healthy tissue (120.8±11Ω, p=0.0015). The mean linear extension of PVGs detected through Lumipoint was significantly lower than the one recognized through voltage map (11.5±8 mm vs 13.3±9 mm, p=0.01) whereas was comparable to the one identified through conventional activation map (11.8±7 mm, p=0.1161 vs Lumipoint). Complete identification of the whole area of PVG was achieved in 31 (62%) and 42 (84%) cases through voltage and activation map, respectively whereas the identification was only partial in 18 (36%) and 7 (14%) cases, respectively. In 1 case both voltage and activation map failed to identify a PVG. No complications during the procedures were reported. All PVs were successfully isolated in all study pts. Conclusion Advanced mapping capabilities were useful to pinpoint the search for PVGs, enabling a more targeted ablation approach vs relying on voltage mapping. LI values correlated well with PVGs characteristics and they significantly differ from both scar and healthy tissue. FUNDunding Acknowledgement Type of funding sources: None.

Automated Mesiodens Classification System Using Deep Learning on Panoramic Radiographs of Children

In this study, we aimed to develop and evaluate the performance of deep-learning models that automatically classify mesiodens in primary or mixed dentition panoramic radiographs. Panoramic radiographs of 550 patients with mesiodens and 550 patients without mesiodens were used. Primary or mixed dentition patients were included. SqueezeNet, ResNet-18, ResNet-101, and Inception-ResNet-V2 were each used to create deep-learning models. The accuracy, precision, recall, and F1 score of ResNet-101 and Inception-ResNet-V2 were higher than 90%. SqueezeNet exhibited relatively inferior results. In addition, we attempted to visualize the models using a class activation map. In images with mesiodens, the deep-learning models focused on the actual locations of the mesiodens in many cases. Deep-learning technologies may help clinicians with insufficient clinical experience in more accurate and faster diagnosis.

Self-Erasing Network for Person Re-Identification

Person re-identification (ReID) plays an important role in intelligent surveillance and receives widespread attention from academics and the industry. Due to extreme changes in viewing angles, some discriminative local regions are suppressed. In addition, the data with similar backgrounds collected by a fixed viewing angle camera will also affect the model’s ability to distinguish a person. Therefore, we need to discover more fine-grained information to form the overall characteristics of each identity. The proposed self-erasing network structure composed of three branches benefits the extraction of global information, the suppression of background noise and the mining of local information. The two self-erasing strategies that we proposed encourage the network to focus on foreground information and strengthen the model’s ability to encode weak features so as to form more effective and richer visual cues of a person. Extensive experiments show that the proposed method is competitive with the advanced methods and achieves state-of-the-art performance on DukeMTMC-ReID and CUHK-03(D) datasets. Furthermore, it can be seen from the activation map that the proposed method is beneficial to spread the attention to the whole body. Both metrics and the activation map validate the effectiveness of our proposed method.