This dissertation addresses a central challenge in gastrointestinal (GI) endoscopic imaging—the inconsistency in image quality, which significantly hampers reliable lesion detection using deep learning–based methods. Unlike other general-purpose i...
This dissertation addresses a central challenge in gastrointestinal (GI) endoscopic imaging—the inconsistency in image quality, which significantly hampers reliable lesion detection using deep learning–based methods. Unlike other general-purpose images, endoscopic imaging is inherently constrained by modality-specific factors such as unpredictable illumination, anatomical artifacts, and inter-operator variation. Endoscopic imaging is inherently susceptible to variations in quality due to diverse environmental and procedural factors, making it challenging to maintain.
This study proposes a framework that identifies and quantifies three fundamental clinical decision error sources in endoscopic imaging: 1) device-oriented artifacts, 2) intrinsic anatomical artifacts, and 3) inter-/intra-operator variation. It introduces targeted strategies to mitigate their respective impacts of each. These sources are addressed through domain-adapted processing strategies suggested by skilled clinicians, aiming to enhance the robustness, consistency, and clinical applicability of deep learning-based lesion detection systems.
Device Oriented Artifacts refer to image quality degradation such as uneven illumination and color channel separation due to rotating light sources. A focus group (skilled clinicians) data review was conducted to address these artifacts. Their insights suggest the design of a domain-adapted image quality assessment (IQA) framework that quantitatively evaluates clarity, surface integrity, and illumination homogeneity. This framework enables reproducible evaluation and mitigation of clinical decision errors arising from device-oriented artifacts. Quantitative evaluations demonstrate that models trained on quality-sorted datasets not only outperform those using randomly sampled or unfiltered data but also achieve faster convergence and improved diagnostic metrics—including sensitivity, specificity, positive predictive value, negative predictive value and accuracy. Comparative assessments against clinician reviews and latency evaluations further support the feasibility of future real-time deployment.
Intrinsic Anatomical Artifacts result from non-pathological intraluminal elements such as seeds, stool, and bubbles, or temporary lens obstructions during endoscopic procedures. As these features are not present in general-purposed datasets, a domain-adapted processing strategy was developed to mitigate their impact. During data curation, such artifacts were explicitly classified and incorporated into training using hard negative mining. This approach enabled the model to distinguish true lesions from non-pathological artifacts, thereby reducing FPs and improving specificity. By embedding clinical domain knowledge into the training process, the model’s reliability in real-world clinical applications was further enhanced.
Inter-/Intra-Operator Variation reflects disparities in adenoma detection rates (ADR) and scope-handling skills among operators, a persistent limitation of endoscopic practice.
Skilled clinicians often rely on contextual cues beyond lesion boundaries, whereas bounding box annotations typically only encode the visible margins, thus failing to fully capture these fine-grained details.
Therefore Soft-margin was incorporated into the YOLOv7 bounding-box regression loss, enabling the model to better accommodate soft lesion boundaries and uncertain margins. This method mirrors the clinical judgment regarding adjacent mucosa and improves model flexibility in boundary interpretation.
Additionally, separate detection models were trained for the upper and lower GI tracts, tailored to their distinct anatomical structures and procedural contexts. An anatomical localization module was integrated to assign detected lesions to standardized GI regions, supporting structured reporting and enhancing interpretability. Altogether, these solutions improve model generalization across variable diagnostic scenarios, enhance robustness to clinical heterogeneity, and strengthen interpretability by linking detections to anatomically meaningful structures.
This work presents a modality-aware, tailored framework for computational GI image analysis by structurally decomposing clinical sources of decision errors and resolving each issue by applying an interpretable, domain-adapted deep learning-based solution—thereby establishing a reproducible and generalizable system for high-accuracy, clinically aligned lesion detection in GI endoscopy.