Development and preclinical evaluation of an endonasal Raman spectroscopy probe for transsphenoidal pituitary adenoma surgery

Abstract

Significance: For most patients with pituitary adenomas, surgical resection represents a viable therapeutic option, particularly in cases with endocrine symptoms or local mass effects. Diagnostic imaging, including MRI and computed tomography, is employed clinically to plan pituitary adenoma surgery. However, these methods cannot provide surgical guidance information in real time to improve resection rates and reduce risks of damage to normal tissue during tumor debulking.

Aim: Here, we present the development of a handheld Raman spectroscopy system that can be seamlessly integrated with transsphenoidal surgery workflows to allow live discrimination of all normal intracranial anatomical structures, including the pituitary gland, and potentially tissue abnormalities such as adenomas.

Approach: A fiber-optic probe was developed with a form factor compatible with endoscopic systems for endonasal surgeries. The instrument was evaluated in an ex vivo experimental protocol designed to assess its ability to distinguish normal intracranial structures. A total of 274 in situ spectroscopic measurements were acquired from six lamb heads, targeting key anatomical structures encountered in surgery. Support vector machine models were developed to classify tissue types based on their spectral signatures.

Results: Binary classification models successfully distinguished the pituitary gland from other tissue structures with a sensitivity and a specificity of 100%. In addition, a four-class predictive model enabled $>95\%$ accuracy in situ discrimination of four structures of most importance during pituitary adenoma tumor resection, i.e., the pituitary gland, the sella turcica (ST) bone, the optic chiasm, and the ST dura mater.

Conclusions: This work sets the stage for the clinical deployment of Raman spectroscopy as an intraoperative real-time decision support system during transsphenoidal surgery, with future work focused on clinical integration and the generalization of the approach to include the detection of tissue abnormalities, such as pituitary adenomas.

Keywords: Raman spectroscopy; biochemistry; machine learning; neurosurgery; pituitary adenoma; tissue optics.

Fig. 1

(a) Handheld endonasal Raman spectroscopy probe. Annotations indicate the different types of sterilizable materials used to fabricate the device. (b) Schematics of the complete system illustrating the enclosed module housing the near-infrared laser, the spectrometer, and the sensor, all connected to the endonasal probe. A control laptop manages all system operations.

Fig. 2

(a) Raw spectroscopic measurement using the near-infrared Raman spectroscopy probe. Repeat measurements were made for each tissue location and averaged to maximize the overall signal-to-noise ratio. A representative measurement is shown for the sella turcica bone. (b) SNV-normalized processed Raman spectrum. (c) Close-up of the endonasal probe positioned at an angle on the sella turcica bone, after removal of the pituitary gland and overlying dura mater.

Fig. 3

Annotated image of a lamb half-head, displaying the various tissues measured using a Raman spectroscopy probe in the scope of the study.

Fig. 4

(a) Spectrogram showing the entire in situ Raman spectroscopy dataset. (b) Average spectra per tissue, with the corresponding inter-measurement variance shown for each spectral bin.

Fig. 5

SAM computed among the average Raman spectra for all possible combinations of tissue types. SAM values range from 0 deg for identical spectral fingerprints to a maximum of 90 deg when the spectra are highly dissimilar.

Fig. 6

Results of the two-class predictive models classifying the pituitary gland and the ST bone. (a) Average Raman spectra with the bands from which features were used for modeling highlighted in blue, confusion matrices representing the classification results associated with the (b) cross-validation and (c) testing phases of model development. (d) Table summarizing the results of all two-class models, including the Raman bands from which features were used by the models, validation and testing predictive accuracies as well as the AUC of the testing set ROC analysis curve.

Fig. 7

Results of the multi-class predictive model (model II). (a) Average Raman spectra of the tissue measurements used to develop a machine learning model. The bands from which features were used for modeling are highlighted in blue. (b) Confusion matrix representing the classification results associated with the cross-validation phase of model development. (c) Confusion matrix showing the results of the machine learning model on the testing hold-out dataset.

Fig. 8

(a) Average Raman spectra of the pituitary glands associated with each tissue preparation method (in situ, ex situ, in-section) with the corresponding inter-measurement variance shown. (b) Variance summed over all spectral bins for each acquisition method and (c) SAM values computed among the average spectra from all combinations of tissue preparation methods. Variances and SAM values are represented for the spectra averaged for each individual lamb head as well as for the spectra averaged across all three lamb heads. Confusion matrices representing the classification results obtained when applying multi-class model II to measurements associated with (d) ex situ and (e) in-section tissue preparation methods.