多型態皮膚軟組織感染傷口之深度學習分割模型

Abstract

皮膚軟組織感染（Skin and Soft Tissue Infection, SSTI）是臨床上常見的感染類型，每年有超過 1400 萬人次因 SSTI 就診、急診或住院，其發生率遠超過肺炎與泌尿道感染，是泌尿道感染的兩倍且是肺炎的十倍。根據美國感染症學會（Infectious Diseases Society of America, IDSA）於 2014 年公布的治療指引，SSTI 的診斷主要依賴醫師的臨床經驗，透過紅、腫、熱、痛等理學檢查觀察病灶。然而，傳統方法存在主觀性高、判讀耗時且難以標準化的問題。在同一觀察者間易出現前後不一致（Intra-rater bias），不同觀察者間亦缺乏一致性（Inter-rater bias）。隨著人工智慧技術的發展，深度學習(Deep learning, DL)開始在自動化診斷皮膚感染中扮演重要角色，可以產生與以往人工判讀或是傳統演算法截然不同的計算方式，不僅為手術後傷口診斷提供了有效的診斷支持，也在其他皮膚病症影像診斷方面已被證實具有高度的效能，能夠識別和分類各種皮膚病變。現今手機攝影的影像品質也已足以應用在臨床案例的量化分析，提供模型大量的影像資料學習以分辨皮膚病變的細微差異，從而判斷 SSTI 感染範圍，提供醫師更快速、客觀及精準的參考資料。 本研究為前瞻性觀察研究，於臺大醫院收集 31 名 SSTI 患者的影像資料，除二位難以分類不納入資料集外，其中 12 位開放性傷口患者共 781 張影像，17 位非開放性傷口患者共 556 張影像，並採用 6-fold 交叉驗證。在開放性傷口分割中，選用 HarDNet-MSEG、HarDNet-DFUS、MFSNet 三種深度學習模型，以深度學習為基礎架構，比較加入預訓練權重，與加入預訓練權重再加上資料擴增組合兩種方法，並在這之上實驗結合 SAM 模型以遮罩點對乘及中心點座標兩種方法進行模型組合測試，探討模型組合的效能。在非開放性（紅腫）傷口分割中，由於邊界不明顯、傷口型態不易判斷等因素任務更為困難，因此採用在開放性傷口分割中表現較為良好的 HarDNet-MSEG、HarDNet-DFUS 兩種深度學習模型進行，比較加入預訓練權重，與加入預訓練權重再加上資料擴增組合兩種方法的效能。並且以 SAM 模型無法分割非開放性傷口，不加入模型組合測試。 結果顯示，於開放性傷口中，以結合預訓練權重並使用資料增強策略，於三個深度學習模型有較高的 Dice，HarDNet-MSEG 為 0.802、HarDNet-DFUS 為 0.795、MFSNet 為 0.722，其中 HarDNet-MSEG 模型達到平均 Dice 分數 0.802，顯著優於其他模型，表現出穩定且高效的分割能力。此外，在 SAM 模型組合成效上，能提升部分影像的分割成效，但在多傷口或形態複雜的傷口影像中受到限制，整體 Dice 分數下降。在非開放性傷口中，同樣以結合預訓練權重並使用資料增強策略，於兩個深度學習模型有較高的 Dice，HarDNet-MSEG 為 0.664、HarDNet-DFUS 為 0.658，同樣以 HarDNet-MSEG 展現最佳效能，其平均 Dice 分數為 0.664，雖低於開放性傷口分割結果，但已證明 HarDNet-MSEG 結合預訓練與資料增強策略有最佳的表現。 整體而言，HarDNet-MSEG 模型在分割任務中均展現了卓越效能與訓練效率，尤其在結合預訓練與資料增強策略時能有效提升分割準確性，開放性傷口達到 0.802 的 Dice，在非開放性（紅腫）傷口達到 0.664 的 Dice，這些研究成果有望在臨床上為 SSTI 患者提供更快速精準的診斷。

Skin and Soft Tissue Infection (SSTI) is one of the most common clinical infections, with over 14 million cases annually in the United States involving outpatient visits, emergency care, or hospitalization. According to the 2014 clinical practice guidelines from the Infectious Diseases Society of America (IDSA), SSTI diagnosis heavily relies on physicians' clinical experience, utilizing physical examination findings such as redness, swelling, heat, and pain to observe and diagnose lesions. However, traditional methods face challenges, including subjectivity, time consumption, and difficulty in standardizing assessments. Even for the same observer, intra-rater bias may arise over time, while inter-rater bias is common across different evaluators, leading to inconsistencies in diagnosis. With advancements in artificial intelligence, deep learning (DL) has begun playing a critical role in the automated diagnosis of skin infections. Deep learning provides an alternative computational approach compared to traditional algorithms or manual interpretation. It has demonstrated exceptional performance not only in diagnosing surgical wound infections but also in identifying and classifying various dermatological conditions through medical imaging. The improved quality of smartphone cameras now enables them to capture clinical-grade images for use in quantitative analysis, providing large datasets for model training. These advancements allow the differentiation of subtle variations in skin conditions and precise delineation of SSTI-infected regions, offering physicians faster, objective, and accurate diagnostic tools. This prospective observational study collected imaging data from 31 SSTI patients at National Taiwan University Hospital. After excluding two unclassifiable cases, the study included 781 images from 12 patients with open wounds and 556 images from 17 patients with non-open wounds. A 6-fold cross-validation approach was employed to ensure the reliability of the data. For open wounds, the study tested three deep learning models—HarDNet-MSEG, HarDNet-DFUS, and MFSNet—using deep learning frameworks and comparing the effects of incorporating pre-trained weights versus pre-trained weights combined with data augmentation. Additionally, two strategies were used to combine the models with SAM: mask multiplication and centroid-based segmentation. For non-open wounds, where boundaries are less distinct and the task is more challenging, the study applied the two best-performing models from the open wound experiments—HarDNet-MSEG and HarDNet-DFUS. The models were evaluated using pre-trained weights alone and with data augmentation. The SAM model was excluded due to its inability to segment non-open wounds. The results showed that for open wounds, combining pre-trained weights with data augmentation achieved higher Dice scores across all three deep learning models. HarDNet-MSEG achieved the highest performance with an average Dice score of 0.802, significantly outperforming the other models (HarDNet-DFUS: 0.795, MFSNet: 0.722). While the SAM combination improved segmentation for some images, its performance declined in cases with multiple or complex wound morphologies, resulting in an overall drop in Dice scores. For non-open wounds, combining pre-trained weights with data augmentation also yielded better results, with HarDNet-MSEG achieving a Dice score of 0.664, outperforming HarDNet-DFUS (0.658). Although the scores were lower than those for open wounds, the findings demonstrated that HarDNet-MSEG combined with pre-training and data augmentation produced the best results. In summary, the HarDNet-MSEG model demonstrated outstanding performance and training efficiency in segmentation tasks, particularly when incorporating pre-trained weights and data augmentation strategies. It achieved Dice scores of 0.802 for open wounds and 0.664 for non-open wounds. These findings highlight the potential of this approach to provide faster and more accurate diagnoses for SSTI patients in clinical settings.