Computer Based Automatic Detection and Classification of Osteoporosis in Bone Radiographs

Machine Learning Solutions for Osteoporosis—A Review

“…( 29,32–34,36–65 ) Osteoporosis classification was made based on lumbar BMD, ( 32–34,37,51 ) hip BMD, ( 38,50,58 ) lumbar and hip BMD, ( 29,39–42,46–48,53,59,60 ) other non‐standard assessments, ( 43,44,49,54–56,65 ) or unspecified. ( 36,45,52,57,61–64 ) Studies identified osteoporosis based on opportunistic imaging from CT, ( 32–34 ) X‐ray, ( 37,38,43–45,55–59,63,64 ) or dental imaging; (36,47–49,53,54,60,62 ) other studies used data from patient characteristics, ( 40,41,50,51,61,65 ) bone biomarkers, (29,39 ) or acoustical responses. ( 42,52 ) As outcome, studies classified osteoporotic versus normal patients, ( 29,36,39,40,43,49,50,52,54–57,62 ) osteoporotic versus non‐osteoporotic patients (based on a BMD T ‐score threshold of –2.5 SD), ( 34,38,44,64 ) normal versus abnormal subjects (based on the BMD T ‐score threshold of −1 SD), ( 33,41,42,45,47,48,58–60,65 ) experimented multiple classifications, ( 46,63 ) or assigned to three classes: osteoporosis (BMD T ‐score ≤ −2.5 SD), osteopenia (−2.5 < BMD T ‐score ≤ −1), and normal (BMD T ‐score > −1 SD).…”

“…Other studies presented a clear risk of overfitting by using data with important bias between case/control groups, (54,58 ) model selection based on the testing set, (29,40,48,51,60 ) reporting high discrepancies between training and test sets, (37,59 ) or including a part of the training data set in the testing data. ( 33,63,65 ) Several studies did not report characteristics of their data set ( 48,55,56,59–64 ) or the model selection process. ( 33,39,41,42,45,46,50,51,53,56–61,64,65 ) Performance was significantly impacted by case prevalence where accuracy dropped from 94.0% to 88.4% when tested on 13% and 50% positive (osteoporotic) cases, respectively.…”

“…( 33,63,65 ) Several studies did not report characteristics of their data set ( 48,55,56,59–64 ) or the model selection process. ( 33,39,41,42,45,46,50,51,53,56–61,64,65 ) Performance was significantly impacted by case prevalence where accuracy dropped from 94.0% to 88.4% when tested on 13% and 50% positive (osteoporotic) cases, respectively. ( 44 ) An image enhancement and standardization step, and combining multiple features, was able to considerably improve results in two studies, respectively.…”

“…( 44 ) An image enhancement and standardization step, and combining multiple features, was able to considerably improve results in two studies, respectively. ( 43,61 ) Overfitting was addressed by cropping images into regions, ( 32,34,37,38,44–46,53,57,59 ) dimensionality reduction to select best features, ( 39,40,43,56,64 ) data augmentation, ( 40,54 ) or by using a previously trained model. ( 38,49,56 ) Curiously, a simple base model using age and body mass index (BMI) did not perform better than random guess (AUC ≈ 0.5), suggesting that the population had undergone a selection process resulting in an imbalance for these variables.…”

“…Based on this checklist, the most critical methodological and reporting limitations observed were presence of overfitting risk, ( 29,33,35,37,40,48,51,54,58–60,63,65 ) lack of model selection configuration reporting, ( 23–25,29,33,35,39,41,42,45,46,50,51,53,56–61,64,65 ) reporting of almost perfect model performances, ( 29,43,48,51,54,56,60,72,92 ) or lack of confidence intervals around point estimates. ( 33,34,44,52–65,72,73,86,88,92,102 ) In general, the studies where these issues were observed were not adequately acknowledged. ( 29,51,60,72,92 ) Acknowledging potential biases or limitations is fundamental for the accurate interpretation and especially for claims related to clinical implication.…”