eess.IVFeb 20, 2026

RamanSeg: Interpretability-driven Deep Learning on Raman Spectra for Cancer Diagnosis

Authors: Chris Tomy, Mo Vali, David Pertzborn, Tammam Alamatouri, Anna Mühlig, Orlando Guntinas-Lichius, Anna Xylander, Eric Michele Fantuzzi, 

Organizations: Department of Computer Science and Technology, University of Cambridge, Cambridge, UK · Department of Physics, University of Cambridge, Cambridge, UK · Department of Otorhinolaryngology, Jena University Hospital, Jena, Germany · Section Pathology of the Institute of Forensic Medicine, Jena University Hospital, Jena, Germany · 

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

Histopathology, the current gold standard for cancer diagnosis, involves the manual examination of tissue samples after chemical staining, a time-consuming process requiring expert analysis. Raman spectroscopy is an alternative, stain-free method of extracting information from samples. Using nnU-Net, we trained a segmentation model on a novel dataset of spatial Raman spectra aligned with tumour annotations, achieving a mean foreground Dice score of 80.9%, surpassing previous work. Furthermore, we propose a novel, interpretable, prototype-based architecture called RamanSeg. RamanSeg classifies pixels based on discovered regions of the training set, generating a segmentation mask. Two variants of RamanSeg allow a trade-off between interpretability and performance: one with prototype projection and another projection-free version. The projection-free RamanSeg outperformed a U-Net baseline with a mean foreground Dice score of 67.3%, offering a meaningful improvement over a black-box training approach.

Figures & tables

Figure 1
Figure 2
Feature Map SizeMean Dice
512×512512\times 512 (Baseline)1.0000
32×3232\times 320.9804
16×1616\times 160.9581
8×88\times 80.9113
4×44\times 40.8356
Table 1: Average dice scores for different bottleneck sizes. The original patch size is 512×512512\times 512 .
Figure 4
ModelDice ( ±\pm std)Sensitivity ( ±\pm std)Specificity ( ±\pm std)
U-Net66.7 ±\pm 15.481.5 ±\pm 23.690.9 ±\pm 5.5
UNet Transformer69.8 ±\pm 10.979.9 ±\pm 18.192.5 ±\pm 3.9
nnU-Net80.9 ±\pm 10.483.5 ±\pm 14.595.9 ±\pm 2.8
RamanSeg60.5 ±\pm 11.795.9 ±\pm 3.479.1 ±\pm 4.9
Projection-free RamanSeg67.3 ±\pm 8.270.3 ±\pm 17.593.3 ±\pm 5.5
Table 5
Figure 6
Figure 7
Figure 8
Figure 9
Figure 10
More in this topic

Image Segmentation

Image segmentation assigns pixels or voxels to objects and regions, supporting measurement, scene understanding, and decisions across image domains.

See all 399 papers

Explore similar work

CardsList

  1. Machine Learning for Biomedical Raman Spectroscopy: From Spectral Acquisition to Clinical Translation

    Jun 12, 2026Bogdan Oancea, Ana Maria Seciu-Grama, Nicoleta Siminea +10Interpretable MLRaman Spectroscopy

    National Institute of Research and Development for Biological Sciences, Romania · University of Bucharest, Romania · University of Turku, Finland +1

  2. Semantic Segmentation-Driven Image-Level Diagnosis of Liver Cancers in Hematoxylin and Eosin Histopathology Images

    Jul 3, 2026Ivica Kopriva, Dario Sitnik, Arijana Pacic +3Computational PathologyHistopathology Image Segmentation

    Division of Computing and Data Science, Ruđer Bošković Institute, Bijenička cesta 54, Zagreb Croatia · Sitnik Ventures UG, Lautensackstr. 25, 80687 Munich, Germany · Department of Pathology and Cytology, Dubrava University Hospital, Avenija Gojka Šuška 6, 10000 Zagreb, Croatia +2

  3. DeepHistoViT: An Interpretable Vision Transformer Framework for Histopathological Cancer Classification

    Mar 12, 2026Ravi Mosalpuri, Mohammed Abdelsamea, Ahmed Karam EldalyHistopathology Image ClassificationComputational Pathology