Advancements in traditional machine learning techniques for detection and diagnosis of fatal cancer types: Comprehensive review of biomedical imaging datasets

Abstract

Accurate cancer detection and diagnosis are imperative for advancing patient outcomes and mitigating mortality rates. This extensive review scrutinizes the progress within the domain of traditional machine learning tech-niques applied to the identification and diagnosis of ten lethal cancer types, utilizing biomedical imaging datasets. The specified cancers encompass lung, colorectal, liver, stomach, breast, esophageal, pancreatic, prostate, brain, and skin cancers. Through a meticulous analysis of 202 literature sources spanning the years 2017 to 2023, this study delves into diverse dimensions of cancer detection and diagnosis. To assess model performance, an array of metrics including accuracy, sensitivity, specificity, precision, negative predictive value, F-measure (F1), area under the curve, Matthews Correlation Coefficient (MCC), Jaccard similarity index, and Youden index is deployed. Notably, the review identifies consistent 100% metric values across various indicators. Furthermore, the abstract provides benchmark and publicly available datasets, accompanied by downloadable links for all the reviewed cancer types, furnishing a valuable repository for prospective research endeavors. The all-encompassing review underscores the remarkable efficacy of traditional machine learning techniques, capable of attaining high accuracy rates, even reaching 100%, in cancer detection tasks despite constraints in datasets and features. This challenges the prevailing orthodoxy that posits deep learning models as universally superior, emphasizing the imperative to delve into and optimize traditional machine learning approaches. The findings posit that further exploration in this realm holds promise for the development of precise and dependable diagnostic tools for cancer.