Colorectal cancer is a major health problem, where advances towards computer-aided diagnosis (CAD) systems to assist the endoscopist can be a promising path to improvement. Here, a deep learning model for real-time polyp detection based on a pre-trained YOLOv3 (You Only Look Once) architecture and complemented with a post-processing step based on an object-tracking algorithm to reduce false positives is reported. The base YOLOv3 network was fine-tuned using a dataset composed of 28,576 images labelled with locations of 941 polyps that will be made public soon. In a frame-based evaluation using isolated images containing polyps, a general F₁ score of 0.88 was achieved (recall = 0.87, precision = 0.89), with lower predictive performance in flat polyps, but higher for sessile, and pedunculated morphologies, as well as with the usage of narrow band imaging, whereas polyp size < 5 mm does not seem to have significant impact. In a polyp-based evaluation using polyp and normal mucosa videos, with a positive criterion defined as the presence of at least one 50-frames-length (window size) segment with a ratio of 75% of frames with predicted bounding boxes (frames positivity), 72.61% of sensitivity (95% CI 68.99–75.95) and 83.04% of specificity (95% CI 76.70–87.92) were achieved (Youden = 0.55, diagnostic odds ratio (DOR) = 12.98). When the positive criterion is less stringent (window size = 25, frames positivity = 50%), sensitivity reaches around 90% (sensitivity = 89.91%, 95% CI 87.20–91.94; specificity = 54.97%, 95% CI 47.49–62.24; Youden = 0.45; DOR = 10.76). The object-tracking algorithm has demonstrated a significant improvement in specificity whereas maintaining sensitivity, as well as a marginal impact on computational performance. These results suggest that the model could be effectively integrated into a CAD system.

Colorectal cancer is one of the most frequent malignancies. Colonoscopy is the de facto standard for precancerous lesion detection in the colon, i.e., polyps, during screening studies or after facultative recommendation. In recent years, artificial intelligence, and especially deep learning techniques such as convolutional neural networks, have been applied to polyp detection and localization in order to develop real-time CADe systems. However, the performance of machine learning models is very sensitive to changes in the nature of the testing instances, especially when trying to reproduce results for totally different datasets to those used for model development, i.e., inter-dataset testing. Here, we report the results of testing of our previously published polyp detection model using ten public colonoscopy image datasets and analyze them in the context of the results of other 20 state-of-the-art publications using the same datasets. The F1-score of our recently published model was 0.88 when evaluated on a private test partition, i.e., intra-dataset testing, but it decayed, on average, by 13.65% when tested on ten public datasets. In the published research, the average intra-dataset F1-score is 0.91, and we observed that it also decays in the inter-dataset setting to an average F1-score of 0.83.

Deep learning object-detection models are being successfully applied to develop computer-aided diagnosis systems for aiding polyp detection during colonoscopies. Here, we evidence the need to include negative samples for both (i) reducing false positives during the polyp-finding phase, by including images with artifacts that may confuse the detection models (e.g., medical instruments, water jets, feces, blood, excessive proximity of the camera to the colon wall, blurred images, etc.) that are usually not included in model development datasets, and (ii) correctly estimating a more realistic performance of the models. By retraining our previously developed YOLOv3-based detection model with a dataset that includes 15% of additional not-polyp images with a variety of artifacts, we were able to generally improve its F1 performance in our internal test datasets (from an average F1 of 0.869 to 0.893), which now include such type of images, as well as in four public datasets that include not-polyp images (from an average F1 of 0.695 to 0.722).

Dysfunctional mitochondria are present in many neurodegenerative diseases, such as spinocerebellar ataxia type 3 (SCA3), also known as Machado–Joseph disease (MJD). SCA3/MJD, the most frequent neurodegenerative ataxia worldwide, is caused by the abnormal expansion of the polyglutamine tract (polyQ) at ataxin-3. This protein is known to deubiquitinate key proteins such as Parkin, which is required for mitophagy. Ataxin-3 also interacts with Beclin1 (essential for initiating autophagosome formation adjacent to mitochondria), as well as with the mitochondrial cristae protein TBK1. To identify other proteins of the mitophagy pathway (according to the KEGG database) that can interact with ataxin-3, here we developed a pipeline for in silico analyses of protein–protein interactions (PPIs), called auto-p2docking. Containerized in Docker, auto-p2docking ensures reproducibility and reduces the number of errors through its simplified configuration. Its architecture consists of 22 modules, here used to develop 12 protocols but that can be specified according to user needs. In this work, we identify 45 mitophagy proteins as putative ataxin-3 interactors (53% are novel), using ataxin-3 interacting regions for validation. Furthermore, we predict that ataxin-3 interactors from both Parkin-independent and -dependent mechanisms are affected by the polyQ expansion.