Class Activation Map
The heat map of CAM image was generated using the otoscopic images from the validation set. The CAM showed the deep learning model can identify attic retraction pocket with red color accurately, and deeper or bigger attic retraction pocket hold more values (Fig 3). Partial atelectasis and general atelectasis were identified by the deep learning model, and deeper or bigger atelectasis showed more values with red color (Fig 4).
DISCUSSION
The diagnosis of attic retraction pocket and atelectasis is based on an otoscopic examination. However, diagnosis of ear diseases with otoscopic images is a hard task for general practitioners, pediatricians and otolaryngologists, with averaged accuracy 39%-53%, 36%-51% and 61%-74% respectively 20-22. In this study, we developed and validated a deep learning model to identify attic retraction pocket and atelectasis with multi-centers otoscopic images. Our CNN algorithm acquired an AUC of 0.89 for the identification of attic retraction pocket and 0.87 for atelectasis.
Previous studies established deep learning models for the diagnosis of tympanic retraction and achieved an averaged accuracy ranging from 85.78% to 88.06% 15,16. Shie et al15 only obtained 856 otoscopic images from one center encompassing almost all otitis media categories. Cha et al16 included 1222 otoscopic images with tympanic retraction and they merged atelectasis and attic retraction pocket into a class. Our study included 6393 OME otoscopic images, of which 55.74% were identified with attic retraction pocket and 38.48% were identified with atelectasis. Considering that the attic retraction pocket was limited to the pars flaccida of tympanic membrane, the atelectasis is in the pars tensa of tympanic membrane. During the progress of disease, attic retraction pocket is more likely to progress to cholesteatoma, and atelectasis is likely to evolve to ossicular erosions10. Therefore, we labeled and divided the retraction pockets into atelectasis and attic retraction pocket. Compared with previous models, we targeted the attic retraction pocket and atelectasis separately, to our current knowledge, this image classification system was the first to diagnose two types of tympanic membrane lesions.
Our results showed different region (pars tensa and pars flaccida) retraction on the tympanic membranes with different accuracy. It is reasonable for clinical experience that attic retraction pocket is easier to identify than atelectasis, because normal tympanic membrane shows a mild retraction in pars tensa without retraction in pars flaccida. The reason may be that in non-severe cases, the attic retraction pocket may be subtle, and clinicians may find it difficult to determine whether this is a normal or a grade I attic retraction pocket based on Tos and Sade classification systems 17,18. On the other hand, cases with severe attic retraction pockets and atelectasis often exposed the ossicles inside the tympanic membrane, and sometimes it is difficult to distinguish between perforation and severe atelectasis.
In order to show the discriminative region of deep learning, CAM highlighted the important area with red color 17, especially large and deep retraction pockets and atelectasis, which was consistent with otologists. Moreover, our image datasets were representative which were collected from three hospitals with different type of otoscopes and image record systems. Many parameters of otoscopes and systems differs in different hospitals, such as the white balance was not equal in different hospital, even in the same hospital because of different preference of practitioners.
During the procedure of follow up, if attic retraction pocket and atelectasis is suggested, observation should be stop and it’s better to triage the patients to otologists. On the other hand, in the clinical practice, this model could be useful for generating diagnosis of attic retraction pocket and atelectasis, which could be assistant for otologists. For young otologists and non-otologists, this model could be used as a study platform to learn attic retraction pocket and atelectasis.
Limitation : Some limitations did exist in our study. Although this CNN algorithm could identify mild and severe attic retraction pocket and atelectasis. However, without enough images of severe attic retraction pocket and atelectasis, it is not easy to develop and validate a deep learning model to identify different level of attic retraction pocket and atelectasis. Moreover, accurate segment labeling techniques may be helpful for improving the accuracy of model. We developed the deep learning model with weak supervision, and further detailed annotation before model development are suggested. Thirdly, non-medical history and hearing information were provided to the deep learning model and otolaryngologists, which may affect the accuracy of diagnosis. The doctor can be greatly improved the accuracy of diagnosis by adding disease history.