Background/Objectives: In a previous study, we utilized categorical variables and machine learning (ML) algorithms to predict the success of non-surgical root canal treatments (NSRCTs) in apical periodontitis (AP), classifying the outcome as either success (healed) or failure (not healed). Given the importance of radiographic imaging in diagnosis, the present study evaluates the efficacy of deep learning (DL) in predicting NSRCT outcomes using two-dimensional (2D) periapical radiographs, comparing its performance with ML models. Methods: The DL model was trained and validated using leave-one-out cross-validation (LOOCV). Its output was incorporated into the set of categorical variables, and the ML study was reproduced using backward stepwise selection (BSS). The chi-square test was applied to assess the association between this new variable and NSRCT outcomes. Finally, after identifying the best-performing method from the ML study reproduction, statistical comparisons were conducted between this method, clinical professionals, and the image-based model using Fisher’s exact test. Results: The association study yielded a p-value of 0.000000127, highlighting the predictive capability of 2D radiographs. After incorporating the DL-based predictive variable, the ML algorithm that demonstrated the best performance was logistic regression (LR), differing from the previous study, where random forest (RF) was the top performer. When comparing the deep learning–logistic regression (DL-LR) model with the clinician’s prognosis (DP), DL-LR showed superior performance with a statistically significant difference (p-value < 0.05) in sensitivity, NPV, and accuracy. The same trend was observed in the DL vs. DP comparison. However, no statistically significant differences were found in the comparisons of RF vs. DL-LR, RF vs. DL, or DL vs. DL-LR. Conclusions: The findings of this study suggest that image-based artificial intelligence models exhibit superior predictive capability compared with those relying exclusively on categorical data. Moreover, they outperform clinician prognosis.

Deep learning applied to chest X-ray (CXR) images has gained wide attention for its potential to improve diagnostic accuracy and accessibility in resource-limited healthcare settings. This study compares two deep learning strategies for lung disease classification: a Two-Stage approach that first detects abnormalities before classifying specific pathologies and a Direct multiclass classification approach. Using a curated database of CXR images covering diverse lung diseases, including COVID-19, pneumonia, pulmonary fibrosis, and tuberculosis, we evaluate the performance of various convolutional neural network architectures, the impact of lung segmentation, and explainability techniques. Our results show that the Two-Stage framework achieves higher diagnostic performance and fewer false positives than the Direct approach. Additionally, we highlight the limitations of segmentation and data augmentation techniques, emphasizing the need for further advancements in explainability and robust model design to support real-world diagnostic applications. Finally, we conduct a complementary evaluation of bone suppression techniques to assess their potential impact on disease classification performance.

Background/Objectives In a previous study, we utilized categorical variables and Machine Learning (ML) algo-rithms to predict the success of non-surgical root canal treatments (NSRCTs) in apical periodontitis (AP), classifying the outcome as either success (healed) or failure (not healed). Given the importance of radiographic imaging in diagnosis, the present study evaluates the efficacy of Deep Learning (DL) in predicting NSRCT outcomes using two-dimensional (2D) periapical radiographs, comparing its performance with ML models. Methods The DL model was trained and validated using Leave-One-Out Cross-Validation (LOOCV). Its output was incorporated into the set of categorical variables, and the ML study was reproduced using Backward Stepwise Selection (BSS). The chi-square test was applied to assess the association between this new variable and NSRCT outcomes. Finally, after identifying the best-performing method from the ML study reproduction, statistical comparisons were conducted between this method, clinical professionals, and the im-age-based model using Fisher’s Exact Test. Results The association study yielded a p-value of 0.000000127, highlighting the predictive ca-pability of 2D radiographs. After incorporating the DL-based predictive variable, the ML algorithm that demonstrated the best performance was Logistic Regression (LR), differing from the previous study, where Random Forest (RF) was the top performer. When comparing the Deep Learning – Logistic Regression (DL-LR) model with the clinician’s prognosis (DP), DL-LR showed superior performance with a statistically significant dif-ference (p-value < 0.05) in sensitivity, NPV, and accuracy. The same trend was observed in the DL vs. DP comparison. However, no statistically significant differences were found in the comparisons of RF vs. DL-LR, RF vs. DL, and DL vs. DL-LR. Conclusions The findings of this study suggest that image-based artificial intelligence models exhibit superior predictive capability compared to those relying exclusively on categorical data. Moreover, they outperform clinician prognosis.

Respiratory diseases remain the third leading cause of death in European Environment Agency member countries, with approximately 420,000 annual deaths and accounting for 6.1% of all EU deaths in 2021 [1,2]. Chest X-rays (CXR), due to their affordability and accessibility, are widely used in diagnostics and screening. However, their interpretation is often hindered by anatomical overlaps and variable pathology presentations, requiring expert radiologists whose availability may be limited in high-demand or resource-constrained environments. Deep learning offers a scalable solution to assist in diagnosis. This study presents a comparative evaluation of two deep learning strategies for CXR-based lung disease classification: (1) a Multiclass approach, in which a single model categorizes CXRs into five classes—four pathologies (pneumonia, COVID-19, tuberculosis, pulmonary fibrosis) and healthy cases; and (2) a Hierarchical approach, where an initial binary model detects abnormality, followed by a secondary model that classifies the specific pathology. This hierarchical strategy is intended to mimic clinical triage, prioritizing abnormality detection before disease differentiation. The inclusion of these four pathologies was motivated by their prevalence, radiographic ambiguity, and clinical relevance in triage settings. Attempts to further subclassify pneumonia into viral and bacterial types using CXRs alone proved unreliable, aligning with known clinical limitations [3] and supporting the selected disease scope. All models were based on DenseNet [4] architectures and trained using a curated, multi-institutional dataset combining over 20,000 images from 10 public and private sources. Images underwent preprocessing steps, including adaptive windowing, lung segmentation using a Weighted Ordered Weighted Averaging (WOWA) ensemble of six segmentation models [5], and targeted data augmentation. A consistent training setup using 5-fold cross-validation was used to ensure robust evaluation. To improve model transparency and clinical trust, we incorporated explainability methods (Grad-CAM [6], Grad-CAM++ [7], and Score-CAM [8]) and verified that activation maps focused on segmented lung regions. Coverage analysis confirmed that models trained with segmentation consistently concentrated their attention within anatomical boundaries, unlike unsegmented models that often relied on irrelevant cues. However, lesion-level validation using bounding boxes revealed limited spatial precision (IoU < 10%), highlighting ongoing challenges in XAI resolution. The Hierarchical approach achieved superior performance in terms of F1 scores for all pathological categories, including Pneumonia (0.9849 vs. 0.9771), COVID-19 (0.9078 vs. 0.9067), Pulmonary Fibrosis (0.9619 vs. 0.9516), and Tuberculosis (0.8926 vs. 0.8786) when compared to the multiclass classification strategy. Additionally, the F1 score for normal cases was also higher (0.9606 vs. 0.9365), indicating more reliable identification of healthy subjects. To evaluate statistical significance, a Shapiro-Wilk test confirmed the normality of the performance data (p > 0.05), enabling the use of a Students t-test. The results show that the Hierarchical Classification approach significantly outperformed the Multiclass Classification method in terms of overall classification performance (p = 0.0392), with a particularly significant improvement in the F1 score (p = 0.0196). A clinically critical metric, the False Positive Rate (FPR) for normal cases, was also notably reduced. The Hierarchical framework achieved a 33% lower FPR for normal cases compared to the multiclass model. This improvement is essential in medical applications, where incorrectly labeling diseased individuals as healthy can delay treatment and endanger patient safety. In summary, the proposed Hierarchical Classification framework demonstrates improved diagnostic performance and interpretability compared to a multiclass model. Its architecture, grounded in clinical logic and validated across multiple datasets, supports its potential for deployment in real-world environments lacking expert radiological oversight.

Type 1 diabetes (T1D) is an autoimmune disease with a strong genetic component (~70% heritability). Early identification of individuals at risk is crucial for early intervention or risk assessment. Although polygenic risk scores (PRS) have shown promise in risk assessment, most current approaches remain constrained by linear assumptions and limited generalizability. We aimed to develop a neural network-driven classifier using T1D-associated single nucleotide polymorphisms (SNPs). In addition, we explored the inclusion of an entropy-derived feature as a complementary variable, representing the degree of genetic variability within an individual’s genotype profile across the 67 T1D-associated SNPs, to evaluate its potential additive contribution to the model performance. We analyzed genotype data from 11,909 individuals in the UK BioBank (546 T1D cases and 11,363 controls). Sixty-seven well-known SNPs associated with T1D were utilized as inputs to the model, using two distinct allele-encoding strategies. A feed-forward neural network was evaluated under varying case–control ratios through five-fold cross-validation. Performance was assessed using the area under the receiver operating characteristic curve (AUC) on a held-out test set and on an external European cohort as a validation cohort. Across five-fold cross-validation, the best configuration achieved a median AUC of 0.903. On the held-out UK Biobank test set, the model generalized well, with an AUC of 0.8889 (95% CI: 0.8516–0.9262). A probability-based risk framework, constructed using five risk groups (“very low”, “low”, “intermediate”, “high”, and “very high” risk), yielded a negative predictive value (NPV) of 98.9% for the “very low” risk group and a Positive Predicted Value (PPV) of 61.9% with a specificity of 97.3% for the “very high” risk group, assuming a 10% T1D prevalence. External validation in the German Diabetes Study reproduced clear case–control separation; for individuals with recent onset diabetes and glutamic acid decarboxylase antibodies (GADA+) vs. controls, specificity reached 91.9% in the “high” risk group (PPV of 94.3%) and 97.6% in the “very high” risk group (PPV of 95.7%). The proposed neural network reliably predicts T1D genetic risk using a compact SNP panel of 67 SNPs and maintains accuracy in both internal and external European cohorts. Its probabilistic output enables clinically interpretable risk thresholds, while entropy features contributed modestly to performance. These results demonstrate that a neural network-based approach achieves discriminative performance that is comparable to established T1D genetic risk models, while offering flexible probability-based risk stratification and architectural extensibility for future integration of additional features.