Spatio-spectral classification of hyperspectral images for brain cancer detection during surgical operations

Abstract

Surgery for brain cancer is a major problem in neurosurgery. The diffuse infiltration into the surrounding normal brain by these tumors makes their accurate identification by the naked eye difficult. Since surgery is the common treatment for brain cancer, an accurate radical resection of the tumor leads to improved survival rates for patients. However, the identification of the tumor boundaries during surgery is challenging. Hyperspectral imaging is a non-contact, non-ionizing and non-invasive technique suitable for medical diagnosis. This study presents the development of a novel classification method taking into account the spatial and spectral characteristics of the hyperspectral images to help neurosurgeons to accurately determine the tumor boundaries in surgical-time during the resection, avoiding excessive excision of normal tissue or unintentionally leaving residual tumor. The algorithm proposed in this study to approach an efficient solution consists of a hybrid framework that combines both supervised and unsupervised machine learning methods. Firstly, a supervised pixel-wise classification using a Support Vector Machine classifier is performed. The generated classification map is spatially homogenized using a one-band representation of the HS cube, employing the Fixed Reference t-Stochastic Neighbors Embedding dimensional reduction algorithm, and performing a K-Nearest Neighbors filtering. The information generated by the supervised stage is combined with a segmentation map obtained via unsupervised clustering employing a Hierarchical K-Means algorithm. The fusion is performed using a majority voting approach that associates each cluster with a certain class. To evaluate the proposed approach, five hyperspectral images of surface of the brain affected by glioblastoma tumor in vivo from five different patients have been used. The final classification maps obtained have been analyzed and validated by specialists. These preliminary results are promising, obtaining an accurate delineation of the tumor area.

Conflict of interest statement

Competing Interests: The authors have declared that no competing interests exist.

Figures

Fig 1. Intra-operative hyperspectral acquisition system used during a neurosurgical procedure at the University Hospital Doctor Negrin of Las Palmas de Gran Canaria.

Fig 2. Screenshot of the HELICoiD Labeling Tool.

Fig 3. Brain cancer detection and delimitation algorithm overview diagram.

(A) HS cube of in-vivo brain surface. (B) Pre-processing stage of the algorithm. (C) Database of labeling samples generation. (D) SVM model training process employing the labeled samples dataset. (E), (F) and (G) Algorithms that conform the spatial-spectral supervised classification stage. (H) and (I) Algorithms that generate the unsupervised segmentation map and the final HELICoiD TMD map, respectively.

Fig 4. Spectral signature of a grade IV glioblastoma tumor tissue.

(A) Raw spectral signature. (B) Calibrated spectral signature. (C) HySIME filtered spectral signature. (D) Final pre-processed spectral signature.

Fig 5. KNN filtered maps obtained with different K and λ values.

(A), (B), (C), (D) and (E) filtered maps obtained with K equal to 5, 10, 20, 40, and 60, while keeping λ value fixed to 1. (F), (G), (H), (I) and (J) filtered maps obtained with λ equal to 0, 1, 5, 10, and 100, while keeping K value fixed to 40.

Fig 6. Hybrid classification example based on a majority voting technique.

The unsupervised segmentation map and the supervised classification maps are merged using the majority voting method.

Fig 7

Mean and variances of the pre-processed spectral signatures of the tumor, normal and blood vessel classes of the labeled pixels from patient 1 (A) and patient 2 (B), represented in red, black and blue color respectively.

Fig 8. Quantitative results of the supervised classification performed with the SVM classifier applied to the labeled data of each patient.

(A) Overall accuracy results of supervised classification per SVM kernel type and patient. (B) and (C) Specificity and sensitivity results obtained using the SVM classifier with linear kernel for each patient and class employing the One-vs-All evaluation method.

Fig 9. Results of each step of the optimized spatial-spectral supervised classification of the five different patients.

(A), (B), (C), (D) and (E) Synthetic RGB images generated from the HS cubes. (F), (G), (H), (I) and (J) Golden standard maps used for the supervised classification training. (K), (L), (M), (N) and (O) Supervised classification maps generated using the SVM algorithm. (P), (Q), (R), (S) and (T) FR-t-SNE one band representation of the HS cubes. (U), (V), (X), (Y) and (Z) Spatially optimized classification maps obtained after the KNN filtering.

Fig 10. Results of each step of the proposed cancer detection algorithm applied to the five different patients.

(A), (B), (C), (D) and (E) Segmentation maps generated using the HKM algorithm. (F), (G), (H), (I) and (J) MV classification maps. (K), (L), (M), (N) and (O) OMD maps that take into account only the major probability per class obtained from the MV algorithm. (P), (Q), (R), (S) and (T) TMD maps that take into account the first three major probabilities per class obtained from the MV algorithm.