Deep learning approach on symptom questionnaire and abdominal radiography for diagnosis of dyssynergic defecation

Abstract

Dyssynergic defecation is one of the most common causes of chronic constipation. It is a behavioral problem in which the pelvic floor muscles are unable to coordinate with the surrounding muscles and nerves to evacuate stool. Patients are required to undergo specialized tests only available�at tertiary healthcare centers for diagnosis. The aim of this thesis is to develop deep learning-based models to prescreen potential patients from primary and secondary healthcare centers for further diagnostic tests by using easily obtainable data such as symptom questionnaire and abdominal radiography. First, we developed a model which uses symptom questionnaire as an input from tree-based machine learning algorithms and deep learning model. Feature selection based on expert knowledge and based on traditional method were performed to find the best set of input features. Second, we developed a model which uses abdominal radiography as an input from the state-of-the-art image classification models. Several image augmentation techniques were�applied as data preprocessing. Third, we proposed an integrated model which uses both symptom questionnaire and abdominal radiography as inputs. The selected input features from symptom questionnaire were combined with image features extracted from the abdominal radiography using a concatenate layer. This approach was meant to imitate how human experts diagnose in real life. We also proposed data preprocessing and postprocessing suitable for small dataset to improve the model accuracy and efficiency. The results show that our proposed integrated model outperforms the baseline models with an accuracy of 66.01%.