Tüberküloz, Mycobacterium tuberculosis bakterisinin neden olduğu, dünya çapında etkili olan ve ölümle sonuçlanabilen bulaşıcı bir hastalıktır. Kişiden alınan balgam örneğinin mikroskop altında incelenmesi hastalığın en yaygın tanı yöntemlerindendir. Ancak bu manuel işlem hataya açık ve zaman alıcıdır. Bu çalışmada, balgam yayma mikroskop görüntüleri kullanılarak otomatik tüberküloz teşhisi için bir model önerilmiştir. Modelde veri setine ait öznitelikler transfer öğrenme yaklaşımı ile Vgg19, ResNet50 ve DenseNet201 kullanılarak elde edilmiştir. Öznitelikler birleştirildikten sonra Komşuluk Bileşen Analizi (KBA) ile ağırlıklandırılıp en iyi olanları seçilmiştir. Seçilen öznitelikler Destek Vektör Makinesi (DVM) ile eğitilmiş, modelin performansı beş kat çapraz doğrulama kullanılarak hesaplanmıştır. Deneysel sonuçlarda %99.03 doğruluk, %99.24 kesinlik, %98.81 duyarlılık, %99.25 özgüllük ve %99.02 f1-skoru elde edilmiştir.

Objective: The aim of this study is to automatically and reliably predict the presence of cardiovascular pathology from heart photographs using deep learning algorithms. Additionally, it aims to compare the diagnostic performance of different deep learning models.Material and Methods: This prospective study was conducted with the approval of the Education and Scientific Research Commission of the Council of Forensic Medicine and the Ethics Committee for Non-Interventional Scientific Research at Tokat Gaziosmanpaşa University. A total of 218 cases (110 men, 108 women) whose hearts were removed during autopsy were included in the study. Photographs of the hearts from each case were obtained using a professional camera. Data augmentation techniques were applied for deep learning analysis, generating six variations from each image and resulting in a dataset comprising a total of 1,308 heart photographs. The hearts were examined in detail both macroscopically and microscopically. Cases with identified cardiovascular pathology were classified as the “patient” group, while cases without any cardiovascular pathology were classified as the “control” group. Morphological images of heart tissue were automatically extracted from the heart photographs. The presence of cardiovascular pathology was assessed using deep learning-based models based on these features.Results: The highest classification accuracy (88.38%) for heart images from the patient and control groups was achieved using the DenseNet201 model. Of the 654 images from the patient group, 569 were correctly classified as “patient,” while 85 were incorrectly predicted as belonging to the “control” group. Conclusion: The study demonstrated that cardiovascular pathology can be predicted using deep learning models based on photographs of the heart’s direct morphological structure, without the need for morphometric measurements. We believe this approach will contribute to forensic reporting processes.

Background/Objectives: The sella turcica is a key anatomical landmark due to its close relationship with the pituitary gland and surrounding structures. This study aimed to compare morphometric characteristics of the sella turcica in autopsy cases according to sex and to develop machine learning (ML)-based sex estimation models using these measurements. Methods: This study included 230 individuals (115 males, 115 females). Sella turcica morphometric measurements (length, depth, anteroposterior, and transverse diameters) were analyzed. In addition, associations with age, height, and weight were evaluated. Sex differences and correlations were assessed using non-parametric tests. Generalized additive models were applied to evaluate non-linear effects of height and weight, and ML algorithms (LR, RF, SVM, XGBoost) were used for sex classification with 10-fold cross-validation. Results: Data from 230 individuals (115 males, 115 females) were analyzed. All sella turcica dimensions were significantly greater in males (p < 0.001). Height showed strong positive correlations with all measurements, whereas weight showed weaker associations and age was not significant. Generalized additive models demonstrated high performance for length and anteroposterior diameter (adjusted R² ≈ 0.956), high performance for transverse diameter (R² = 0.936), and moderate performance for depth (R² = 0.700), with significant non-linear effects of height and weight. ML models achieved high accuracy (>95%), with SVM performing best (accuracy: 0.991; AUC: 0.997), and transverse diameter identified as the most important predictor. Conclusions: Sella turcica morphometry demonstrates strong sexual dimorphism and is primarily influenced by body size parameters, particularly height. Non-linear modelling approaches such as GAM effectively capture complex anatomical relationships, while ML models, especially SVM, provide promising sex estimation. Among all variables, transverse diameter emerges as the most robust and consistent predictor, highlighting its potential utility in forensic and anthropological applications.