Gallbladder cancer (GBC) is globally rare but prevalent in India. GBC is often diagnosed at an advanced stage, leading to a poor outcome. The identification of key prognostic factors and accurate survival prediction are crucial to optimizing treatment strategies. Retrospective data from 698 patients with gallbladder cancer treated at Tata Medical Center, Kolkata were obtained from a published dataset. Different machine learning models were used to predict survival outcomes. The performance of the model was evaluated, and explainable AI techniques were used to interpret the model’s output and identify significant prognostic factors. The analysis identified elevated liver enzymes and bilirubin levels as significant prognostic factors. Advanced age has also been shown to be correlated with terminal disease outcomes. When comparing different modeling systems, the stacking model obtained perfect matrix scores. These results confirmed the resilience of the stacking model for survival prediction. Other models also showed high prediction performance. Explainable AI techniques provide details of the relative importance of these prognostic factors. Compared to prevailing models of survival prediction for gallbladder cancer, this one represents a major improvement, correlating a rigorously validated stacking ensemble with comprehensive explainability, as demonstrated by SHAP, PDP, ICE, and LIME. This results in almost perfect discrimination (AUROC = 0.9949) and a significantly boosted interpretability of prognostic factors. The partial dependence plots present the effects of liver and glycemic markers on survival prognosis. SHAP values illustrated that the final stage of the tumor and the absence of surgery are the top negative prognostic indicators. The explainability of SHAP also demonstrated the importance of liver enzymes, bilirubin, age, and albumin levels in predicting survival. Hence, explainable machine learning models help predict more accurately and also support an understanding of the main factors contributing to survival. If they are ever integrated into clinical environments, such models can help personalize treatment strategies and improve care outcomes. All codes and datasets used in this study are available at https://github.com/devnarayan87/GBC_Cancer_Data .
Vasculature feature segmentation from the fundus images is critical for the identification of retinal diseases. However, automated vessel segmentation is challenging owing to variability in vessel structure and color gradients, low contrast between vessels, and pathologies. The state-of-the-art approaches address the challenges of emerging hand-crafted filters to apprehend vessel- like patterns. More recently, deep learning-based methods have evolved for the vessel segmentation task, which employs annotations to train the model. These approaches ignore the vessel's geometrical characteristics in the fundus images, leading to inaccuracies. Herein, we propose a novel unsupervised segmentation method based on interrelationship handling, Bonferroni mean pre-aggregation operator, with the aid of dynamic fuzzy histogram equalization, namely BMPDFHESeg. The method extracts the vascular information by fusing color channels by constructing an interrelationship handling pre-aggregation operator. The operator enables finding the direction of increasingness to segment large vessels and vessel feature enhancement through a dynamic fuzzy histogram equalization process using the prior feature intensity information. BMPDFHESeg is assessed qualitatively and quantitatively using the DRIVE, STARE, and HRF datasets, demonstrating enhanced efficacy and computational speed. Further, the results were validated by ophthalmologists for the accuracy of the vessel segmentation and usefulness for the diagnosis of retinal disorders.
In this article, we propose the generalized version of the extended, partitioned Bonferroni mean () operator with a systematic investigation of its behavior and properties. It can aggregate data of various dimensions in one formulation by modeling mandatory conditions along with partitioned structure interrelationships amongst the criterion set. In addition, we generate the condition for weight vectors satisfied by the weighting triangle associated with the proposed extended aggregation operator. We employed the proposed operator to aggregate a dataset following a hierarchical structure. We found that by implementing the proposed operator one can even rank the alternatives more intuitively with respect to any intermediate perspective of the hierarchical system. Finally, we present an application of the proposed extended aggregation operator in a case-based example of a child’s home environment quality evaluation with detailed analysis.