Scholarlink

<p>As a core tool for minimally invasive diagnosis and treatment of gastrointestinal, respiratory, and urinary diseases, endoscopes are now used in increasingly precise surgeries. AI-assisted endoscopic sequence depth estimation has thus become a key need, as it provides 3D spatial guidance to improve safety. To address the limitations of traditional methods in the task of endoscopic depth estimation, this paper proposes an innovative framework comprising a Dual-Path Feature Aggregation Pyramid Module (DPFAP) and a Deep Feature-Aware Reconstruction Module (DFAR), where DPFAP is the primary structural innovation and DFAR is the optimization strategy for the loss function. Most studies rely on image pyramids and feature pyramids; while the former exhibit strong semantics, they lack interaction between hierarchical information, and although the latter enable cross-level information transmission, they suffer from poor semantic consistency. Thus, this paper proposes a hierarchical alignment and fusion strategy (DPFAP) specifically for the image pyramid and the feature pyramid mentioned above. Furthermore, traditional unsupervised methods rely on the “Lambertian surface” assumption and are prone to interference from endoscopic environments, which in turn leads to compromised training processes and poor reconstruction performance. To address this, the DFAR module is introduced, which adopts an innovative feature reconstruction loss strategy: It synthesizes middle-layer features and calculates the loss based on these features, using the result as additional supervision. On the SCARED endoscopic depth estimation dataset, the proposed algorithm achieves state-of-the-art (SOTA) performance with an Abs Rel of 0.049, an Sq Rel of 0.321, an RMSE of 4.265, an RMSE log of 0.065, and a δ of 0.984. Compared with current SOTA counterparts, the Abs Rel and RMSE are reduced by ~ 3.9% and ~ 1.9%, respectively, while the δ is increased by 0.3%, with all core depth estimation accuracy metrics significantly improved. Furthermore, generalization experiments on the Hamlyn endoscopic dataset demonstrate that the proposed algorithm possesses excellent scene adaptability and generalization capability, enabling its effective application to various endoscopic depth estimation scenarios, which validates the effectiveness and practicality of the innovative designs in this work.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Self-Supervised Endoscopic Depth Estimation via Deep Feature-Aware Reconstruction and Dual-Path Feature Aggregation Pyramid

  • 期刊论文

  • 发表日期：

    2026年3月4日

查看全文
  • Yukang Ren,
  • Yanping Chen

摘要

As a core tool for minimally invasive diagnosis and treatment of gastrointestinal, respiratory, and urinary diseases, endoscopes are now used in increasingly precise surgeries. AI-assisted endoscopic sequence depth estimation has thus become a key need, as it provides 3D spatial guidance to improve safety. To address the limitations of traditional methods in the task of endoscopic depth estimation, this paper proposes an innovative framework comprising a Dual-Path Feature Aggregation Pyramid Module (DPFAP) and a Deep Feature-Aware Reconstruction Module (DFAR), where DPFAP is the primary structural innovation and DFAR is the optimization strategy for the loss function. Most studies rely on image pyramids and feature pyramids; while the former exhibit strong semantics, they lack interaction between hierarchical information, and although the latter enable cross-level information transmission, they suffer from poor semantic consistency. Thus, this paper proposes a hierarchical alignment and fusion strategy (DPFAP) specifically for the image pyramid and the feature pyramid mentioned above. Furthermore, traditional unsupervised methods rely on the “Lambertian surface” assumption and are prone to interference from endoscopic environments, which in turn leads to compromised training processes and poor reconstruction performance. To address this, the DFAR module is introduced, which adopts an innovative feature reconstruction loss strategy: It synthesizes middle-layer features and calculates the loss based on these features, using the result as additional supervision. On the SCARED endoscopic depth estimation dataset, the proposed algorithm achieves state-of-the-art (SOTA) performance with an Abs Rel of 0.049, an Sq Rel of 0.321, an RMSE of 4.265, an RMSE log of 0.065, and a δ of 0.984. Compared with current SOTA counterparts, the Abs Rel and RMSE are reduced by ~ 3.9% and ~ 1.9%, respectively, while the δ is increased by 0.3%, with all core depth estimation accuracy metrics significantly improved. Furthermore, generalization experiments on the Hamlyn endoscopic dataset demonstrate that the proposed algorithm possesses excellent scene adaptability and generalization capability, enabling its effective application to various endoscopic depth estimation scenarios, which validates the effectiveness and practicality of the innovative designs in this work.