A deep-learning approach to aid in diagnosing Barrett’s oesophagus related dysplasia

Barrett's oesophagus is the only known precursor to oesophagus carcinoma. Histologically, it is defined as a condition of columnar cells replacing the standard squamous lining. Those altered cells are prone to cytological and architectural abnormalities, known as dysplasia. The dysplastic degree varies from low to high grade and can evolve into invasive carcinoma or adenocarcinoma. Thus, detecting high-grade and intramucosal carcinoma during the surveillance of Barrett's oesophagus patients is vital so they can be treated by surgical resection. Unfortunately, the achieved interobserver agreement for grading dysplasia among pathologists is only fair to moderate. Nowadays, grading Barrett's dysplasia is limited to visual examination by pathologists for glass or virtual slides. This work aims to diagnose different grades of dysplasia in Barrett’s oesophagus, particularly high-grade dysplasia, from virtual histopathological slides of oesophagus tissue.
In the first approach, virtual slides were analysed at a low magnification to detect regions of interest and predict the grade of dysplasia based on the analysis of the virtual slides at 10X magnification. Transfer learning was employed to partially fine-tune two deep-learning networks using healthy and Barrett’s oesophagus tissue. Then, the two networks were connected. The proposed model achieved 0.57 sensitivity, 0.79 specificity and moderate agreement with a pathologist.
On the contrary, the second approach processed the slides at a higher magnification (40X magnification). It adapted novelty detection and local outlier factor alongside transfer learning to solve the multiple instances learning problem. It increased the performance of the diagnosis to 0.84 sensitivity and 0.92 specificity, and the interobserver agreement reached a substantial level.
Finally, the last approach mimics the pathologists’ procedure to diagnose dysplasia, relying on both magnifications. Thus, their behaviours during the assessment were analysed. As a result, it was found that employing a multi-scale approach to detect dysplastic tissue using a low magnification level (10X magnification) and grade dysplasia at a higher level (40X magnification). The proposed computer-aided diagnosis system was built using networks from the first two approaches. It scored 0.90 sensitivity, 0.94 specificity and a substantial agreement with the pathologist and a moderate agreement with the other expert.