Comparison of different classifiers to recognize active bone marrow from CT images

One of the main problems during in the treatment of anal cancer with chemotherapy and radiation is the occurrence of Hematologic Toxicity (HT). In particular, during radiotherapy it is crucial to spare Bone Marrow (BM), since the radiation dose received by BM in pelvic bones predicts the onset of HT. In this direction, the most popular strategies are based on the identification of the hematopoietically active BM (actBM), that is the part of BM in charge of blood cells generation, using MRI, SPECT or PET, but no approached have been proposed based on CT. In this study we compare four different classifiers in recognizing actBM from CT images using 36 radiomic features. We used Genetic Algorithms (GAs) to simultaneously optimize the feature subsets and the classifier parameters, separately for three pelvic subregions: iliac bone marrow (IBM), lower pelvis bone marrow (LPBM), and lumbosacral bone marrow (LSBM). The obtained classifiers were applied to CT sequences of a cohort of 25 patients affected by carcinoma of the anal canal. Classifiers results were compared with the actBM identified from 18FDG-PET (reference standard, RS). It emerged that the performances of the 4 classifiers are similar and they are satisfactory for IBM and LSBM subregions (Dice > 0.7) whereas they are poor for LPBM (Dice < 0.5).