Automatic International Hidradenitis Suppurativa Severity Score System (AIHS4): A novel tool to assess the severity of hidradenitis suppurativa using artificial intelligence.

BACKGROUND: Hidradenitis suppurativa (HS) is a painful chronic inflammatory skin disease that affects up to 4% of the European adult population. International Hidradenitis Suppurativa Severity Score System (IHS4) is a dynamic scoring tool that was developed to be incorporated into the doctor's daily clinical practice and clinical studies. This helps measure disease severity and guides the therapeutic strategy. However, IHS4 assessment is a time-consuming and manual process, with high inter-observer variability and high dependence on the observer's expertise. MATERIALS AND METHODS: We introduce the Automatic International Hidradenitis Suppurativa Severity Score System (AIHS4), an automatic equivalent of IHS4 that deploys a deep learning model for lesion detection, called Legit.Health-IHS4net, based on the YOLOv5 architecture. AIHS4 was trained on Legit.Health-HS-IHS4, a collection of HS images manually annotated by six specialists and processed by a novel knowledge unification algorithm. RESULTS: Our results show that, with the current dataset size, our tool assesses the severity of HS cases with a performance comparable to that of the most expert physician. Furthermore, the model can be implemented into CADx systems to support doctors in their clinical practice and act as a new endpoint in clinical trials. CONCLUSION: Our work proves the potential usefulness of artificial intelligence in the practice of evidence-based dermatology: models trained on the consensus of large clinical boards have the potential to empower dermatologists in their daily practice and replace current standard clinical endpoints.

Machine Learning Algorithms as a Computer-Assisted Decision Tool for Oral Cancer Prognosis and Management Decisions: A Systematic Review.

INTRODUCTION: Despite multiple prognostic indicators described for oral cavity squamous cell carcinoma (OCSCC), its management still continues to be a matter of debate. Machine learning is a subset of artificial intelligence that enables computers to learn from historical data, gather insights, and make predictions about new data using the model learned. Therefore, it can be a potential tool in the field of head and neck cancer. METHODS: We conducted a systematic review. RESULTS: A total of 81 manuscripts were revised, and 46 studies met the inclusion criteria. Of these, 38 were excluded for the following reasons: use of a classical statistical method (N = 16), nonspecific for OCSCC (N = 15), and not being related to OCSCC survival (N = 7). In total, 8 studies were included in the final analysis. CONCLUSIONS: ML has the potential to significantly advance research in the field of OCSCC. Advantages are related to the use and training of ML models because of their capability to continue training continuously when more data become available. Future ML research will allow us to improve and democratize the application of algorithms to improve the prediction of cancer prognosis and its management worldwide.

Machine Learning Models for Predicting Facial Nerve Palsy in Parotid Gland Surgery for Benign Tumors.

BACKGROUND: Despite the increasing use of intraoperative facial nerve monitoring during parotid gland surgery (PGS) and the improvement in the preoperative radiological assessment, facial nerve injury (FNI) remains the most severe complication after PGS. Until now, no studies have been published regarding the application of machine learning (ML) for predicting FNI after PGS. We hypothesize that ML would improve the prediction of patients at risk. METHODS: Patients who underwent PGS for benign tumors between June 2010 and June 2019 were included. RESULTS: Regarding prediction accuracy and performance of each ML algorithm, the K-nearest neighbor and the random forest achieved the highest sensitivity, specificity, positive predictive value, negative predictive value F-score, receiver operating characteristic (ROC)-area under the ROC curve, and accuracy globally. The K-nearest neighbor algorithm achieved performance values above 0.9 for specificity, negative predictive value, F-score and ROC-area under the ROC curve, and the highest sensitivity and positive predictive value. CONCLUSIONS: This study demonstrates that ML prediction models can provide evidence-based predictions about the risk of FNI to otolaryngologists and patients. It is hoped that such algorithms, which use clinical, radiological, histological, and cytological information, can improve the information given to patients before surgery so that they can be better informed of any potential complications.

Automatic Urticaria Activity Score: Deep Learning-Based Automatic Hive Counting for Urticaria Severity Assessment.

Chronic urticaria is a chronic skin disease that affects up to 1% of the general population worldwide, with chronic spontaneous urticaria accounting for more than two-thirds of all chronic urticaria cases. The Urticaria Activity Score (UAS) is a dynamic severity assessment tool that can be incorporated into daily clinical practice, as well as clinical trials for treatments. The UAS helps in measuring disease severity and guiding the therapeutic strategy. However, UAS assessment is a time-consuming and manual process, with high interobserver variability and high dependence on the observer. To tackle this issue, we introduce Automatic UAS, an automatic equivalent of UAS that deploys a deep learning, lesion-detecting model called Legit.Health-UAS-HiveNet. Our results show that our model assesses the severity of chronic urticaria cases with a performance comparable to that of expert physicians. Furthermore, the model can be implemented into CADx systems to support doctors in their clinical practice and act as a new end point in clinical trials. This proves the usefulness of artificial intelligence in the practice of evidence-based medicine; models trained on the consensus of large clinical boards have the potential of empowering clinicians in their daily practice and replacing current standard clinical end points in clinical trials.

Facial nerve palsy following parotid gland surgery: A machine learning prediction outcome approach.

Introduction: Machine learning (ML)-based facial nerve injury (FNI) forecasting grounded on multicentric data has not been released up to now. Three distinct ML models, random forest (RF), K-nearest neighbor, and artificial neural network (ANN), for the prediction of FNI were evaluated in this mode. Methods: A retrospective, longitudinal, multicentric study was performed, including patients who went through parotid gland surgery for benign tumors at three different university hospitals. Results: Seven hundred and thirty-six patients were included. The most compelling aspects related to risk escalation of FNI were as follows: (1) location, in the mid-portion of the gland, near to or above the main trunk of the facial nerve and at the top part, over the frontal or the orbital branch of the facial nerve; (2) tumor volume in the anteroposterior axis; (3) the necessity to simultaneously dissect more than one level; and (4) the requirement of an extended resection compared to a lesser extended resection. By contrast, in accordance with the ML analysis, the size of the tumor (>3 cm), as well as gender and age did not result in a determining favor in relation to the risk of FNI. Discussion: The findings of this research conclude that ML models such as RF and ANN may serve evidence-based predictions from multicentric data regarding the risk of FNI. Conclusion: Along with the advent of ML technology, an improvement of the information regarding the potential risks of FNI associated with patients before each procedure may be achieved with the implementation of clinical, radiological, histological, and/or cytological data.

Automatic SCOring of Atopic Dermatitis Using Deep Learning: A Pilot Study.

Atopic dermatitis (AD) is a chronic, itchy skin condition that affects 15-20% of children but may occur at any age. It is estimated that 16.5 million US adults (7.3%) have AD that initially began at age >2 years, with nearly 40% affected by moderate or severe disease. Therefore, a quantitative measurement that tracks the evolution of AD severity could be extremely useful in assessing patient evolution and therapeutic efficacy. Currently, SCOring Atopic Dermatitis (SCORAD) is the most frequently used measurement tool in clinical practice. However, SCORAD has the following disadvantages: (i) time consuming-calculating SCORAD usually takes about 7-10 minutes per patient, which poses a heavy burden on dermatologists and (ii) inconsistency-owing to the complexity of SCORAD calculation, even well-trained dermatologists could give different scores for the same case. In this study, we introduce the Automatic SCORAD, an automatic version of the SCORAD that deploys state-of-the-art convolutional neural networks that measure AD severity by analyzing skin lesion images. Overall, we have shown that Automatic SCORAD may prove to be a rapid and objective alternative method for the automatic assessment of AD, achieving results comparable with those of human expert assessment while reducing interobserver variability.

Constellation Loss: Improving the Efficiency of Deep Metric Learning Loss Functions for the Optimal Embedding of histopathological images.

BACKGROUND: Deep learning diagnostic algorithms are proving comparable results with human experts in a wide variety of tasks, and they still require a huge amount of well-annotated data for training, which is often non affordable. Metric learning techniques have allowed a reduction in the required annotated data allowing few-shot learning over deep learning architectures. AIMS AND OBJECTIVES: In this work, we analyze the state-of-the-art loss functions such as triplet loss, contrastive loss, and multi-class N-pair loss for the visual embedding extraction of hematoxylin and eosin (H&E) microscopy images and we propose a novel constellation loss function that takes advantage of the visual distances of the embeddings of the negative samples and thus, performing a regularization that increases the quality of the extracted embeddings. MATERIALS AND METHODS: To this end, we employed the public H&E imaging dataset from the University Medical Center Mannheim (Germany) that contains tissue samples from low-grade and high-grade primary tumors of digitalized colorectal cancer tissue slides. These samples are divided into eight different textures (1. tumour epithelium, 2. simple stroma, 3. complex stroma, 4. immune cells, 5. debris and mucus, 6. mucosal glands, 7. adipose tissue and 8. background,). The dataset was divided randomly into train and test splits and the training split was used to train a classifier to distinguish among the different textures with just 20 training images. The process was repeated 10 times for each loss function. Performance was compared both for cluster compactness and for classification accuracy on separating the aforementioned textures. RESULTS: Our results show that the proposed loss function outperforms the other methods by obtaining more compact clusters (Davis-Boulding: 1.41 ± 0.08, Silhouette: 0.37 ± 0.02) and better classification capabilities (accuracy: 85.0 ± 0.6) over H and E microscopy images. We demonstrate that the proposed constellation loss can be successfully used in the medical domain in situations of data scarcity.

Artificial Neural Network as a Tool to Predict Facial Nerve Palsy in Parotid Gland Surgery for Benign Tumors.

(1) Background: Despite the increasing use of intraoperative facial nerve monitoring during parotid gland surgery or the improvement in the preoperative radiological assessment, facial nerve injury (FNI) continues to be the most feared complication; (2) Methods: patients who underwent parotid gland surgery for benign tumors between June 2010 and June 2019 were included in this study aiming to make a proof of concept about the reliability of an artificial neural networks (AAN) algorithm for prediction of FNI and compared with a multivariate linear regression (MLR); (3) Results: Concerning prediction accuracy and performance, the ANN achieved the highest sensitivity (86.53% vs 46.23%), specificity (95.67% vs 92.59%), PPV (87.28% vs 66.94%), NPV (95.68% vs 83.37%), ROC-AUC (0.960 vs 0.769) and accuracy (93.42 vs 80.42) than MLR; and (4) Conclusions: ANN prediction models can be useful for otolaryngologists-head and neck surgeons-and patients to provide evidence-based predictions about the risk of FNI. As an advantage, the possibility to develop a calculator using clinical, radiological and histological or cytological information can improve our ability to generate patients counselling before surgery.