We validated and published a landmark paper 1, 2 describing a breakthrough two-stage Convolutional Neural Network (CNN) based deep learning algorithm trained, validated, and tested on cervical images selected from a dataset of 9,450-woman, population-based longitudinal cohort (ages 18-92) acquired by the NCI in Guanacaste, Costa Rica. The cohort provided cervical training/validation images and clinical endpoints. The algorithm, called automated visual evaluation (AVE) of the acetowhitened cervix, generates a severity score (0 to 1) that indicates the likelihood that the cervix in the image is precancerous. Applied to enrollment cervical images, it outperformed standard screening tests (clinician interpretation of the same cervical images, Pap smears, and even HPV testing) in predicting cumulative risk of precancer/cancer. AVE provides sensitive screening with minimal clinical training or cost. Overtreatment still must be addressed by further improvements, unless lower sensitivity is accepted. The AVE algorithm depends on good quality images where quality is a combination of optimal lighting, adequate framing of the cervix, absent any occluding artefacts (e.g., swab, speculum, unrelated anatomy, etc.), and sharp focus. Of these, sharp focus is relatively easily measurable and subsumes lighting and cervix framing. We developed two novel deep learning algorithms one that detects if a cervix is in focus 3 and the other that uses adversarial deep learning networks to deblur out of focus images 4. Both algorithms can be further engineered to be a part of image acquisition phase that would provide acceptable quality images to the AVE algorithm. Cytology / Pap smear analysis is the non-inferiority test for AVE. We have started work toward creating a novel deep learning algorithm-based cytology image classifier for whole-slide images. We worked on a small deidentified dataset of liquid pap-smear slides from Beckton-Dickinson that came from a joint study they participated in with NCI. Our deep learning algorithm operated on cytologist marked high-sensitive regions on the whole slide image; i.e. those that contained high likelihood of abnormal cells, to detect and segment nuclei and classify them as one of several abnormal categories. For this we transferred knowledge from another cervical cytology slide dataset which only provided truth as individual segmented cells. Other cervicographic images, we also continued efforts toward furthering prior work in deep learning-based classification of histopathology images. We continue to develop a novel algorithm that localizes epithelial region of interest on the image to which prior work in image classification can then be applied. In prior efforts, the region of interest was manually marked.
| Sornapudi, Sudhir; Stanley, Ronald Joe; Stoecker, William V et al. (2018) Deep Learning Nuclei Detection in Digitized Histology Images by Superpixels. J Pathol Inform 9:5 |
