Kopriva Ivica, Sitnik Dario, Dion-Bertrand Laura-Isabelle, Periša Marija Milković, Pačić Arijana, Hadžija Mirko, Hadžija Marijana Popović
Division of Computing and Data Sciences, Ruđer Bošković Institute, Bijenička cesta 54, Zagreb, Croatia.
School of Computation, Information Technology (CIT), Technical University Munich, Arcisstraße 21, 80333, Munich, Germany.
Comput Biol Med. 2025 Sep;196(Pt B):110841. doi: 10.1016/j.compbiomed.2025.110841. Epub 2025 Jul 25.
Hyperspectral imaging (HSI) holds significant potential for transforming the field of computational pathology. However, the number of HSI-based research studies remains limited, and in many cases, the advantages of HSI over traditional RGB imaging have not been conclusively demonstrated, particularly for specimens collected intraoperatively. To address these challenges we present: (i) a database consisted of 27 HSIs of hematoxylin-eosin stained frozen sections, collected from 14 patients with colon adenocarcinoma metastasized to the liver. It is aimed to validate pixel-wise classification for intraoperative tumor resection; (ii) a novel method which combines Grassmann points with nearest subspace classifier for pixel-wise classification of HSIs. The HSIs were acquired in the spectral range of 450 nm-800 nm, with a resolution of 1 nm, resulting in images of 1384 × 1035 pixels. Pixel-wise annotations were performed by two pathologists and one medical expert. To overcome challenges such as experimental variability and the lack of annotated data, we applied Grassmann manifold (GM) approach in combination with spectral-spatial features extracted by tensor singular spectrum analysis (TSSA) method to non-overlapping patches of 230 × 258 pixels. Using only 1 % of labeled pixels per class, the GM-TSSA method achieved a micro balanced accuracy (BACC) of 0.963 and a micro F-score of 0.959 on the HSI dataset. The GM-TSSA approach outperformed six deep learning architectures trained with 63 % of labeled pixels. Data are available at: https://data.fulir.irb.hr/islandora/object/irb:538, and code is available at: https://github.com/ikopriva/ColonCancerHSI.
高光谱成像（HSI）在变革计算病理学领域方面具有巨大潜力。然而，基于HSI的研究数量仍然有限，并且在许多情况下，HSI相对于传统RGB成像的优势尚未得到确凿证明，特别是对于术中采集的标本。为应对这些挑战，我们提出：（i）一个数据库，由27张苏木精-伊红染色冰冻切片的高光谱图像组成，这些图像取自14例发生肝转移的结肠腺癌患者。其目的是验证术中肿瘤切除的逐像素分类；（ii）一种将格拉斯曼点与最近子空间分类器相结合的新方法，用于高光谱图像的逐像素分类。高光谱图像在450纳米至800纳米的光谱范围内采集，分辨率为1纳米，生成1384×1035像素的图像。逐像素注释由两名病理学家和一名医学专家进行。为克服实验变异性和注释数据缺乏等挑战，我们将格拉斯曼流形（GM）方法与通过张量奇异谱分析（TSSA）方法提取的光谱-空间特征相结合，应用于230×258像素的非重叠图像块。GM-TSSA方法在高光谱图像数据集上仅使用每个类别1%的标记像素，就实现了0.963的微观平衡准确率（BACC）和0.959的微观F分数。GM-TSSA方法优于使用63%标记像素训练的六种深度学习架构。数据可在以下网址获取：https://data.fulir.irb.hr/islandora/object/irb:538，代码可在以下网址获取：https://github.com/ikopriva/ColonCancerHSI。
