Roman Voitetskii

Background: Breast cancer is one of the leading causes of female mortality, especially if diagnosed in late stages. While mammography is the cornerstone of screening, its diagnostic accuracy is limited by tumor heterogeneity and subjective interpretation. Objective: Herein, we explored the potential of radiomics and machine learning to improve the diagnostic accuracy of mammograms and personalise patient management in breast cancer. Methods: We manually segmented tumours and lymph nodes to analyse mammograms of the open-source INbreast dataset, which comprised multiple cases of benign and malignant breast masses with and without lymphadenopathy. Ra- diomics features (morphological, texture, wavelet) were extracted using PyRadiomics. Stratified sampling ensured balanced class representation. Then, we trained ML classifiers (XGBoost, CatBoost, LightGBM, etc.) to detect malignancy from the extracted radiomical features. Random Forest classifier was used to prognosticate the molecular subtype of the tumour from radiomical findings. Results: Significant radiomic differences were observed between benign and malignant lesions. Combining features of breast mass and lymph node yielded the highest classification accuracy (up to 99%) in detecting malignancy. The Random Forest model achieved 90.8% accuracy in identifying Luminal A molecular subtypes, with first-order and shape-based features contributing most to model perfor- mance. Conclusion: Radiomics-based ML models significantly improve diagnostic accuracy and enable non-invasive prediction of breast cancer subtypes. This approach supports precision oncology by enhancing screening efficiency and informing personalized treatment strategies.

Background and Objective: Corneal collagen cross‐linking (CXL) is a treatment which arrests keratoconus (KC) progression, but its effectiveness differs radically among patients. Herein, we report preoperative diagnostic findings that reflect CXL outcomes and allow physicians to prognosticate treatment efficiency. Methods: In a medical centre, we retrospectively analysed pre‐ and postoperative data about 107 patients (112 eyes) treated with CXL from January 2018 to December 2022. Exclusion criteria were age below 16 years, a corneal thickness below 400 microns, severe dry eye, other corneal diseases/infections, re‐CXL, pregnancy and missing follow‐up examinations. All the subjects (79 males and 28 females) were followed for a minimum of 4 and a maximum of 40 months. The study dataset was comprised of 796 cases of clinical assessment, pachymetry, visiometry, refractometry and topography examinations. With these data, we modelled maximum anterior keratometry ( K _max ) and curvature power of the flat and steep meridians of the corneal anterior surface ( K ₁ and K ₂ ). Results: Two years after the invasion, corneal curvature coefficients decreased progressively. Then, they remained stable for four months and rose afterwards. In the most accurate K ₁ , K ₂ and K _max models, the proportion of mean absolute error to the range of values was 1.72, 3.66 and 2.37%, respectively. Pronounced corneal thinning, low best‐corrected visual acuity and high K _max levels predict unfavourable outcomes. Conclusions: The high accuracy of the models advocates for a personalised approach to candidate selection for CXL.