Machine learning for evolutive lymphoma and residual masses recognition in whole body diffusion weighted magnetic resonance images

The application of machine learning to imaging in hematological oncology: A scoping review

“…In studies that compared different ML methods on the same datasets ( 30 , 32 , 34 , 35 , 42 , 43 , 50 , 51 , 55 ), no single approach consistently outperformed the others.…”

“…Of the 33 studies that applied ML techniques to diagnose a hematological malignancy or to differentiate it from another disease state or malignancy ( Table 1 ), 18 were designed to establish and train ML models to discriminate gliomas [predominantly GBM from PCNSL ( 29 – 31 , 34 – 37 , 40 – 42 , 44 , 45 , 49 , 50 , 53 – 56 )] using features extracted from FDG-PET [one study ( 29 ),] or MRI ( 30 , 31 , 34 – 37 , 40 – 42 , 44 , 45 , 49 , 50 , 53 – 56 ) images. The remaining studies belonged to two major categories: those developing models to discriminate solid hematological malignancies from other benign and malignant lesions at other sites [nasopharyngeal carcinomas from nasopharyngeal lymphoma ( 46 , 48 ), idiopathic orbital inflammation from ocular adnexal lymphoma ( 33 ), thymic neoplasm from thymic lymphoma ( 14 ), breast carcinoma from breast lymphoma ( 15 ), lymphoma from normal nodes ( 43 ), or multiple myeloma from bone metastases ( 51 )] and those that detect the location of hematological malignancies either at diagnosis or during the disease course [location of ( 18 ) or evolving/residual lymphoma ( 32 ) or leukemia ( 17 ) or bone marrow involvement with multiple myeloma ( 16 , 38 , 47 , 52 ) or mantle cell lymphoma ( 39 )].…”

“…Many ML techniques were used, and, in some studies, different modeling approaches were compared on the same dataset ( 30 , 32 , 34 , 35 , 42 , 43 , 50 , 51 , 55 ). Others developed models using a single approach or a combination of approaches in an end-to-end manner ( 31 , 32 , 55 , 56 ).…”

“…Many ML techniques were used, and, in some studies, different modeling approaches were compared on the same dataset ( 30 , 32 , 34 , 35 , 42 , 43 , 50 , 51 , 55 ). Others developed models using a single approach or a combination of approaches in an end-to-end manner ( 31 , 32 , 55 , 56 ). The following diverse ML approaches were used to discriminate lymphomas from other benign or malignant lesions: support vector machines (SVMs ( 29 – 31 , 33 – 37 , 46 , 48 , 50 , 51 , 53 – 55 );), linear discriminant analysis (LDA ( 14 , 15 , 30 , 34 , 37 );), logistic regression (LR ( 30 );), artificial/convolutional neural networks (A/CNNs ( 31 , 40 , 45 , 49 , 51 , 55 , 56 );), k -nearest neighbors (K-NNs ( 34 , 51 );), naïve Bayes classification (NB ( 34 , 50 , 51 );), decision trees (DTs ( 34 );), random forests (RFs ( 34 , 35 , 43 , 44 , 50 , 51 , 55 );), adaptive boosting ( 34 ), and gradient boosting ( 41 , 43 ).…”

“…The following diverse ML approaches were used to discriminate lymphomas from other benign or malignant lesions: support vector machines (SVMs ( 29 – 31 , 33 – 37 , 46 , 48 , 50 , 51 , 53 – 55 );), linear discriminant analysis (LDA ( 14 , 15 , 30 , 34 , 37 );), logistic regression (LR ( 30 );), artificial/convolutional neural networks (A/CNNs ( 31 , 40 , 45 , 49 , 51 , 55 , 56 );), k -nearest neighbors (K-NNs ( 34 , 51 );), naïve Bayes classification (NB ( 34 , 50 , 51 );), decision trees (DTs ( 34 );), random forests (RFs ( 34 , 35 , 43 , 44 , 50 , 51 , 55 );), adaptive boosting ( 34 ), and gradient boosting ( 41 , 43 ). The ML approaches used to detect the location of hematological malignancies either at diagnosis or during the course of disease were similarly diverse: A/CNNs ( 18 , 32 , 48 , 77 ), SVMs ( 32 , 38 ), K-NN ( 32 , 38 ), RF ( 16 , 17 , 32 ).…”