AI-Enhanced Hyperspectral Imaging for Intraoperative Detection of Low-Grade Glioma: A Clinical Study

1. Introduction

• Distinguishing LGG from healthy tissue is challenging because LGGs can closely resemble healthy tissue, while current neuronavigation methods face accuracy, cost, complexity, interpretability, or fluorescence limitations.Preoperative MRI accuracy decreases with brain shift; intraoperative MRI is costly and complex, ultrasound can be difficult to interpret, and most LGGs do not fluoresce.Source: §1 · p. 2
• HSI captures many contiguous spectral bands alongside spatial information, and prior intraoperative brain HSI studies have mainly examined high-grade glioma, glioblastoma, or mixed-grade cohorts.Prior work includes pixel-wise classification using random forests, CNNs, and hybrid spatial-spectral frameworks.Source: §1 · pp. 2–3
• The clinical study enrolled 18 patients and retained 17 for model development, focusing on low-grade glioma rather than the high-grade, glioblastoma, or mixed-grade cohorts prevalent in prior intraoperative HSI studies.One patient was excluded because fluorescence imaging was used.Source: §1 · p. 3
• The study compares its LGG dataset with public HELiCoiD data to assess inter-patient and intra-patient spectral variability across tissue types.The comparison contributes to standardized assessment of spectral imaging for surgical guidance, particularly for LGG.Source: §1 · pp. 3–4
• HSI and RGB models are compared on the same tissue regions using identical patient-level splits and a common evaluation procedure.This paired design is intended to assess the contribution of HSI information under matched conditions.Source: §1 · p. 4

2.1.1. Hardware setup and data acquisition

• The internal system paired a microscope-mounted hyperspectral camera with RGB imaging, using a reference target for calibration and sequential RGB-then-HSI acquisition of the same region.Images of the brain surface were acquired before, during, and after resection.Source: §2.1.1 · pp. 4–5

2.1.2. Data preprocessing

• After dark correction, white balancing, spectral correction, and band removal, the internal images contained 104 contiguous bands spanning 470–787 nm.RGB images and delineations were registered to the HSI frame using target-based calibration followed by rigid registration; hypercubes were cropped to 800 × 800 pixels.Source: §2.1.2 · pp. 6–7

2.1.3. Ground truth processing

• Ground-truth processing retained histology-confirmed healthy and LGG superpixels and removed mixed-label superpixels, intensity outliers, and high-variability regions.Intensity values outside each image’s 5th–95th percentile range were excluded; superpixels above image-specific median thresholds for spectral-shape variability and SAM were also removed.Source: §2.1.3 · pp. 7–8
• Superpixel spectral-shape filtering used mean L2 distance from normalized pixel spectra to the superpixel mean, retaining lower-variability regions.Superpixels above the image-specific 50th percentile of this distance were excluded; smaller values indicate greater within-superpixel spectral-shape homogeneity.Source: §2.1.3 · p. 9

2. Data

• HELiCoiD provides a public multi-grade comparison dataset, while the clinical study enrolled 18 patients and acquired 75 brain images; 17 patients remained for modeling and 16 contributed to patient-level results.Patient 10 was excluded because fluorescence imaging was used, and patient 3 had only healthy annotations and was excluded from binary evaluation.Source: §2 · p. 4, §2.1.1 · pp. 5–6, §2.2 · p. 10

2.2. HELiCoiD dataset (public)

• The public HELiCoiD dataset contains 61 images from 34 patients across LGG, HGG, GBM, and other brain tumors, recorded with 826 channels spanning 400–1000 nm.Its setup differs from the internal imaging system.Source: §2.2 · p. 10

3.1. Spectral characterization of glioma datasets

• The study characterizes spectral properties and their variation across tissue types, datasets, and patients to inform brain-tumor classification for surgical resection.The cross-dataset comparison uses 100 matched bands from each dataset and calculates SAM after pairing internal bands with the closest HELiCoiD wavelengths.Source: §3.1 · p. 10, §3.1.1 · p. 10

3.1.2. Inter-patient versus intra-patient variability

• Intra-patient variability is the mean angular distance of pixel spectra from the corresponding patient-level mean spectrum, calculated separately by tissue class and dataset.Pixels are equally weighted within a patient, and patient-level values are equally weighted in dataset-level distributions.Source: §3.1.2 · p. 11
• Inter-patient variability averages spectral distances across unique patient pairs, comparing patient-level mean spectra with equal weight for each pair.Higher values indicate greater spectral variation between patients.Source: §3.1.2 · p. 11, Figure 3 · p. 12

3.2. Classification methods

• The spectral XGBoost classifier uses each superpixel’s normalized mean spectrum, while CNNs process neighboring spectra to incorporate local spatial information.Evaluated CNN designs include 2D models using all channels or compressed channels and a 3D CNN combining spatial and spectral features.Source: §3.2 · pp. 12–13, Figure 4 · p. 13

4.1. Study design

• Six outer patient-level folds each serve once as test; the other five are used for inner tuning, while patient 3 is excluded from outer testing.All images and superpixels from a patient remain in the same fold, and the five inner-fold models’ predictions are averaged for the outer test.Source: §4.1 · p. 14, Figure 5 · p. 15

4.2. Experiments

• Several HSI machine-learning and deep-learning architectures are benchmarked against a 2D CNN trained on RGB images using the same patient-level splits.Source: §4.2 · p. 14, §4.1 · p. 14

4.3. Training details

• Models train for 50 epochs with binary cross-entropy and Adam, while balanced sampling selects 100 superpixels per class and patient each epoch.Source: §4.3 · p. 14

4.4. Evaluation

• Performance is measured with six metrics, and each HSI model is compared with RGB using paired patient-level Wilcoxon signed-rank tests.ROC-AUC uses continuous probabilities; the other metrics classify probabilities at a 0.5 threshold.Source: §4.4 · p. 16

5.1. Spectral characterization

• Tumor tissue had lower reflectance than healthy tissue across both datasets, with similar spectral features despite different HSI systems.The spectra were similar near 584 nm; the authors note that hemoglobin-related absorption may contribute to shared features.Source: §5.1 · p. 16, Figure 6 · p. 17
• Tumor tissue had higher intra-patient SAM in both datasets, significantly so only in InVivo-LGG, and higher inter-patient SAM across datasets (p < 0.05).Source: §5.1 · pp. 16–17, Figure 7 · p. 18

5.2. Classification results

• XGBoost had the highest median F1, accuracy, and recall; 3D CNN had the highest balanced accuracy and ROC-AUC, and 2DCNN-ch3 had the highest precision.Source: §5.2 · p. 17
• HSI-based models performed significantly better overall than RGB, while no statistically significant differences appeared among HSI model families.Table 2 summarizes per-patient test-set metrics as medians and 95% confidence intervals.Source: Table 2 · p. 17, §5.2 · p. 17
• HSI models achieved higher patient-level F1 and ROC-AUC than RGB for most patients in the outer test folds.Source: §5.2 · p. 19, Figure 8 · p. 19

• The study adds a clinical low-grade glioma cohort and a paired HSI–RGB comparison using matched tissue locations, identical patient-level splits, and common evaluation.Reflectance-curve shapes also shared features across datasets acquired with different clinical HSI systems.Source: §6 · p. 19
• Tumor tissue had lower reflectance, but proposed links to water content, edema, and vascularity remain hypotheses because these properties were not directly measured.Source: §6 · p. 19
• The HELiCoiD 720-nm trough’s possible hypoxia interpretation is unvalidated, and its absence from InVivo-LGG may reflect physiological or acquisition differences.No physiological measurements were available to validate the interpretation.Source: §6 · p. 21
• Differences in instruments, resolution, illumination, acquisition protocols, and ground-truth procedures complicate direct comparison between the datasets.Source: §6 · p. 21
• Ground-truth procedures differ: HELiCoiD uses semi-automatic spectral labeling, while this study broadly delineates regions using preoperative imaging and selected histology.The authors identify accurate intraoperative registration and spatially matched histopathology, especially at tumor margins, as an ideal.Source: §6 · pp. 21–22
• No HSI architecture was uniformly superior; transformers were deferred because the cohort lacks the training data they generally require.The HSI-over-RGB result is specific to the selected tissue regions and cohort studied.Source: §6 · p. 22

7. Limitations

• A single surgeon provided the tumor and healthy-tissue ground truth, with no inter-rater or intra-rater agreement measurements.The task also included only healthy and tumor tissue, excluding vessels and other background classes relevant to clinical practice.Source: §7 · pp. 22–23
• RGB–HSI sample selection and filtering were based on HSI, which may favor HSI models despite evaluation at matched locations.The authors recommend ablations without spectral filtering or modality-independent evaluation regions to quantify this selection bias.Source: §7 · p. 23
• HSI-based filtering excluded mixed-label boundaries and other common intraoperative conditions, so results apply to selected homogeneous regions, not general tumor detection or reliable margin delineation.Excluded conditions included vessels, blood contamination, specular reflections, very dark regions, and spectrally heterogeneous tissue.Source: §7 · p. 23
• Filtering may alter estimated reflectance distributions and reduce measured within-class variability, while retention at each filtering step was not quantified.The authors recommend reporting stepwise retention and repeating spectral comparisons before and after filtering.Source: §7 · p. 23
• Neural-network spatial context is limited because classification operates on superpixels rather than segmentation that includes local and global spatial information.The authors identify broader segmentation-based approaches as future work.Source: §7 · p. 24

8. Conclusion

• 74.5% median ROC-AUC and 63.7% median F1 for the 3D CNN, compared with 64.0% and 55.6%, respectively, for the RGB-based model.The 3D CNN evaluates spectral signatures within a local spatial context.Source: §8 · p. 24
• The study focuses on low-grade glioma, complementing prior intraoperative HSI studies centered on high-grade glioma, glioblastoma, or mixed-grade cohorts.It compares HSI and RGB using matched tissue locations and identical patient-level evaluation splits.Source: §8 · p. 24
• Both HSI systems show similar brain-tissue spectral features despite system differences, while tumor tissue has greater inter-patient spectral variability than healthy tissue.Source: §8 · p. 24

Appendix A. Registration Examples

• Figure A.12 shows examples of good and bad RGB-to-HSI registrations; bad registrations were manually registered afterward.Source: Figure A.12 · p. 32

Appendix B. Data characteristics

• Table B.3 lists all 18 enrolled patients and their image and superpixel counts before and after filtering; patient 10 was excluded from model development and evaluation.Patient 3 had only healthy-tissue annotations, so it was used for training but excluded from patient-level binary classification evaluation.Source: Table B.3 · p. 33

Appendix C. Binary classification metrics

• Table C.4 is a binary confusion matrix, and accuracy is the proportion of correct predictions among all predictions.Source: Table C.4 · p. 34, Appendix C · p. 34
• Precision is the proportion of positive predictions that are true positives, while recall is the proportion of actual positives correctly identified.Source: Appendix C · p. 34
• Specificity is the proportion of actual negatives correctly identified, and F1 is the harmonic mean of precision and recall.Source: Appendix C · p. 34
• The appendix identifies ROC-AUC as area under the receiver-operating-characteristic curve and notes that false-positive rate varies with the decision threshold.It denotes the false-positive rate as u.Source: Appendix C · p. 34

Appendix D. Results cross-validations

• Figure D.13 visualizes the cross-validation results of hyperparameter optimization for each run.Each boxplot sample represents a run’s cross-validation score for a hyperparameter combination.Source: Appendix D · p. 35, Figure D.13 · p. 36