Computer Vision Identifies Recurrent and Nonrecurrent Ductal Carcinoma in Situ Lesions with Special Emphasis on African-American Women

What Applied Physical Chemistry Can Contribute to Understanding Cancer: Toward the Next Generation of Breakthroughs

“…Computational analysis of glycolysis has shown that its key rate-limiting steps are glucose transport, hexokinase action, PFK catalysis, and lactic acid export from cells [89]. Glucose uptake via phospho-Ser-226-GLUT1 and PFK1 (PFKL) steps were found to be of especial interest in retrospective case-control clinical trials, as they serve as a prognostic test for cancer recurrences [90][91][92][93]. In recurrent cancer cells, these two biomarkers underwent trafficking to the plasma membrane (and elsewhere), presumably to enhance glucose uptake.…”

“…Immunofluorescence micrographs of the two proteins described above, phospho-Ser226-GLUT1 and PFKL (Figure 4), are sufficient to accurately predict the outcomes (recurrences or non-recurrences) of DCIS patients [92,93]. To establish prognoses at a high level of accuracy, we employed computer vision.…”

“…Pathologists observe several variables in micrographs, such as color and size, whereas computer vision analyzes many millions of variables to find several dozens of mathematical variables that predict recurrences and non-recurrences. Over the past few decades, pathologists have been asked to do more, including predicting disease outcomes, and this is now possible using machine tools to assist pathologists [93].…”

“…Training datasets were immunofluorescence micrographs of patient samples stained with antiphospho-Ser226-GLUT1 and PFKL. Therefore, one could perform binary classifications of patients with unknown outcomes [38,72,[90][91][92][93]. For example, the computer was trained with images such as those of Figure 4.…”

“…Other locations of this antigen were also learned by the computer including blood vessels, cancer-associated fibroblasts, myoepithelial cells in combination with appropriate weighting factors to arrive at a predicted outcome. In this way, a patient's outcome could be predicted many years in advance at high accuracy for patients held out of the training dataset [93]. The software identified recurrent cases with no false negatives, non-recurrent women, and false positives due to the efficacy of mastectomies [93].…”