Addressing Challenges in Deep Learning-Based Diagnosis for Pancreatic Cancer and Bile Duct Obstruction

Abstract

In this comprehensive study, two distinct challenges in medical imaging diagnosis were addressed using deep learning techniques. Part 1: Invasive pancreatic cancer diagnosis in endoscopic ultrasound images leveraging translation models. In guiding treatment decisions for pancreatic cancer, assessing vascular invasion is critical, particularly for determining resectability. Deep learning techniques have demonstrated potential for diagnosing vascular invasion through the analysis of pancreatic endoscopic ultrasound (EUS) images. However, challenges arise when dealing with multicenter data sources and imbalanced datasets, which may affect the model performance. EUS images were collected from 170 patients with diagnosed pancreatic cancer at three endoscopy centers using various ultrasound systems. To diagnose vascular invasion while mitigating data variation, feature and image translation models were utilized to effectively align the source and target domains. An image translation model was utilized in the proposed approach (multicenter transfer learning (MCTL)) by employing CycleGAN and customized weighted loss classification models. The performance was compared with a feature translation model (multicenter domain adaptation (MCDA)) and widely accepted baseline classification models. The translation models compensated for distinctive data-specific features and improved the models for classifying vascular invasion. Although the feature translation model proved effective, its applicability was limited across different datasets. The proposed MCTL approach showed superior classification performance with accuracy improvements of 26.79%, 67.26%, and 50.91% over the baseline model and 17.86%, 48.81%, and 42.50% over the MCDA model for the three different imbalanced datasets. This research utilized a deep learning methodology to improve the identification of vascular invasion in pancreatic cancer through the analysis of EUS images from multiple centers and addressed the issue of data imbalance. Part 2: Multimodal deep learning-based diagnosis of extrahepatic common bile duct obstruction using MRCP images and clinical parameters. Magnetic resonance cholangiopancreatography (MRCP) is a noninvasive technique used for the diagnosis of extrahepatic common bile duct obstruction (EHBDO). However, its interpretation is laborious. This study aimed to devise a classification pipeline that automatically assists radiologists in diagnosing EHBDO by utilizing multimodal data such as MRCP images and clinical parameters. This study included 465 patients. Object detection was employed to enhance the common bile duct (CBD) region from MRCP images. Clinical parameters were collected from patients, compensated for missing values by imputation, and subjected to statistical tests. Using a single modality (clinical parameters or MRCP images), machine and deep learning approaches evaluated the diagnostic capabilities of each modality for detecting obstructions. To diagnose CBD obstruction and malignancy, a deep learning approach combining features from both MRCP images and clinical parameters was proposed. A weighted loss function addressing variations in importance of features was applied to each modality. The proposed deep learning model demonstrated superior performance compared to the use of a single modality. It achieved notable overall accuracy, AUC score, sensitivity, and specificity rates of 89.8%, 90.4%, 81.8%, 95.7%, for detecting obstruction and 77.0%, 82.5%, 67.2%, 91.2%, for identification of malignancy, respectively. Overall, this study showcases the potential of deep learning methodologies to improve diagnostic accuracy and efficiency in medical imaging.;본 연구에서는 딥러닝 기술을 사용하여 의료 영상 진단의 두 가지 과제를 해결했다. 첫 번째로 변환 모델을 활용한 내시경 초음파 영상의 췌장암 진단 방법을 연구했다. 췌장암 치료 결정을 내릴 때 혈관 침범을 평가하는 것은 수술에서의 절제 가능성을 결정하는 데 중요하다. 딥러닝 기술로 췌장 내시경 초음파(EUS) 영상 분석을 통해 혈관 침범을 진단할 수 있는 가능성을 입증했다. 연구에 사용한 데이터는 다중 센터에서 획득한 불균형 데이터 세트로 딥러닝 모델의 성능에 큰 영향을 미치는 것을 확인하였다. 3개 내시경 센터에서 진단된 췌장암 환자로부터 다양한 초음파 시스템을 이용하여 EUS 영상을 수집하고 혈관 침범을 진단하기 위해 특성 및 이미지 변환 모델을 활용하여 소스 및 대상 도메인을 효과적으로 변환했다. 제안된 multicenter translation (MCTL) 접근 방식은 세 가지 서로 다른 불균형 데이터 세트에 대해 우수한 분류 성능을 보여주었다. 본 연구에서는 딥러닝 접근 방식을 활용하여 다중 센터의 EUS 이미지에서 췌장암의 혈관 침범 진단을 기능을 향상시키고 데이터 불균형 문제를 해결했다. 두 번째로 자기공명 췌담관조영술(MRCP) 영상과 임상 지표를 이용한 다중 모드 딥러닝 기반 간외 총담관 폐쇄 진단에 대한 연구이다. 자기공명 췌담관조영술은 간외 총담관 폐쇄(EHBDO) 진단에 사용되는 비침습적 기술이다. 본 연구에서는 MRCP 영상과 임상 지표의 다중 모드 데이터를 활용하여 EHBDO 진단을 자동으로 지원하는 분류 파이프라인을 고안하는 것을 목표로 했다. 총담관(CBD) 폐쇄 및 악성종양을 진단하기 위해 MRCP 이미지와 임상 매개변수의 특징을 결합한 딥러닝 접근 방식이 제안되었다. 단일 모드 데이터(임상 매개변수 또는 MRCP 이미지)를 사용한 머신러닝 및 딥러닝 접근 방식과 진단 기능을 비교했을 때 제안된 다중 모드 데이터를 활용한 딥러닝 모델은 우수한 성능을 보였다.