Forecasting Endodontic Outcomes with Deep Convolution Neural Network by Analyzing Intraoral Radiographs

Abstract

Forecasting Endodontic Outcomes with Deep Convolution Neural Network by Analyzing Intraoral Radiographs Junghwan Hwang, D.D.S. Department of Dentistry The Graduate School, Yonsei University (Directed by Professor Sunil Kim, D.D.S., M.S.D., Ph.D.) There have been efforts to develop endodontic prognosis prediction models using artificial intelligence, and PRESSAN-17 was presented. PRESSAN-17 is a deep convolution neural network (DCNN) model that detects clinical features and predicts endodontic outcomes based on preoperative periapical radiographs. However, this model had some limitations, especially using a small dataset while training, validating, and testing. This study aimed to validate and generalize the DCNN model using a larger dataset to increase the reliability of the model, generalize it to cover all types of teeth, and improve its overall performance. A database of all types of teeth that received endodontic treatment or retreatment by endodontists with the presence of a one-year outcome was prepared (n=2,603). There were some alterations to the PRESSAN-17’s structure, replacing the average pool with an attention pool and concatenating clinical features to prediction. This updated model which named ApexNet-17 was trained, validated, and tested to predict the one-year endodontic prognosis by analyzing preoperative periapical radiographs and comparing them to the result of the PRESSAN-17. Accuracy and sensitivity were mainly compared, and gradient-weighted class activation mapping was used to visualize weighted heatmaps. The ApexNet-17 had an accuracy of 72% and a sensitivity of 65.8%. This result emphasizes the model's enhanced ability to predict treatment failures. The basis of the model’s prediction was checked by using a superimposed heatmap on preoperative periapical radiographs by Grad-CAM. Despite some limitations, deep convolutional neural networks have demonstrated their probability to predict outcomes using clinical features identified in periapical radiographs effectively. Their integration with dental imaging can offer an opportunity for enhanced diagnostic accuracy and support more informed clinical decision-making. Efforts have been made to develop models for predicting endodontic prognoses using artificial intelligence, with PRESSAN-17 being one such model. PRESSAN-17 is a deep convolutional neural network (DCNN)-based model that identifies clinical features and forecasts endodontic outcomes from preoperative periapical radiographs. However, the model faced certain limitations, notably the use of a small dataset during its training, validation, and testing phases. This study aimed to validate and enhance the DCNN model by employing a larger dataset, thereby increasing the model's reliability, extending its applicability to all tooth types, and improving its overall performance. A database comprising all types of teeth that underwent endodontic treatment or retreatment by endodontists, with a one-year outcome recorded, was established (n=2,603). Modifications were made to the structure of PRESSAN-17, which included replacing the average pool with an attention pool and appending clinical features to enhance prediction. The revised model, named ApexNet-17, was trained, validated, and tested to predict the one-year endodontic prognosis by analyzing preoperative periapical radiographs. Its performance was compared with that of PRESSAN-17, focusing primarily on accuracy and sensitivity. Additionally, gradient-weighted class activation mapping was employed to generate weighted heatmaps for visualization. ApexNet-17 achieved an accuracy of 72% and a sensitivity of 65.8%, indicating the model's improved capability to predict treatment failure. The basis of the model's predictions was verified through the application of a heatmap superimposed on preoperative periapical radiographs using Grad-CAM. Despite certain limitations, deep convolutional neural networks have shown potential to effectively predict outcomes by utilizing clinical features identified in periapical radiographs. The incorporation of these networks into dental imaging could provide an opportunity to improve diagnostic accuracy and facilitate more informed clinical decision-making.

구내 방사선 영상 분석을 통한 심층 컨볼루션 신경망의 근관치료 결과 예측 황 정 환 연세대학교 대학원 치의학과 (지도교수 김 선 일) 수술 전 치근단 방사선 사진을 기반으로 임상적 특징을 감지하고 근관치료 결과를 예측하는 것으로 제안된 심층 컨볼루션 신경망 모델인 PRESSAN-17은 작은 데이터셋을 기반으로 했다는 한계점이 있었다. 이 연구는 더 큰 데이터 세트를 사용하여 심층 컨볼루션 신경망 모델을 검증하고 일반화하는 것을 목표로 모델의 신뢰성을 높이고, 모든 유형의 치아를 적용할 수 있도록 일반화하며, 전반적인 성능을 개선하기 위해 수행되었다. 연세대학교 치과대학병원 치과보존과 혹은 연세대학교 강남 세브란스 치과보존과에서 보존과 전문의에게 근관치료 또는 재근관치료를 받았으며 1년간의 경과 관찰을 진행한 모든 유형의 치아에 대한 2,603개의 데이터베이스를 구성하였다. 또한PRESSAN-17에서 약간의 구조 수정이 있었으며 이는 average pooling을 attention pooling으로 변경한 것과 임상적 특징을 예후 예측 과정에 통합하는 것이었다. 변형된 모델은 ApexNet-17로 명명하였으며 수술 전 치근단 방사선 사진을 분석하여 1년 근관 치료 예후를 예측하도록 훈련, 검증 및 테스트하고 PRESSAN-17의 결과와 비교했다. 정확도와 민감도를 주로 비교했으며, 가중 히트맵(weighted heatmaps)을 시각화하기 위해 기울기 가중 클래스 활성화 매핑(Grad-CAM)을 사용했다. ApexNet-17의 정확도는 이전 연구의 67.0%에서 72.0%로 증가했다. 민감도의 경우 40.7%에서 65.8%로 개선되어 치료 실패를 예측하는 모델의 성능이 향상됨을 보였다. 정밀도, F1점수, 수신자 조작 특성 곡선 밑의 면적 역시 크게 향상된 것으로 나타났다. Grad-CAM의 수술 전 치근단 방사선 사진의 Grad-CAM을 통해 모델 예측의 근거를 확인할 수 있다. 몇 가지 한계가 있지만 심층 컨볼루션 신경망은 치근단 방사선 사진에서 확인된 임상적 특징을 사용하여 결과를 효과적으로 예측할 수 있는 가능성을 입증했다. 심층 컨볼루션 신경망 모델과 치과 영상과의 통합은 진단 정확도를 높이고 보다 정보에 입각한 임상 의사 결정을 지원할 수 있는 기회를 제공할 수 있을 것으로 보인다.