An adaptive stacking generalization integrated with Raman spectroscopy feature enhancement algorithm for fine glioma grading identification.

For different grades of brain gliomas, it is crucial for clinicians to rapidly and accurately develop personalized treatment strategies intraoperatively to improve surgical outcomes and enhance the quality of life for patients. Raman fiber miniature spectroscopy detection can provide detailed information about the properties of biomolecules. This technique offers several advantages, including non-invasiveness, real-time detection, intelligence, high precision, and the potential for early diagnosis. Therefore, it facilitates the development of portable, low-cost, and non-invasive in situ and in vivo tumor grading devices. However, distinguishing between low-grade and high-grade gliomas is challenging due to minimal grade differences. The low signal-to-noise ratio inherent to Raman fiber miniature spectrometers also result in subtle spectral features. These factors pose significant difficulties for conventional recognition algorithms. To address this issue, an innovative fine target recognition algorithm based on adaptive stacking generalization integrated with Raman spectroscopy feature enhancement (ASG-RSFE) is proposed in this paper. Unlike traditional methods that directly utilize Raman spectral data for modeling, this paper proposes a Raman characteristic peak ratio approach for feature enhancement. This method effectively amplifies subtle biomolecular changes induced by glioma lesions in the brain. Additionally, this study introduces the butterfly optimization algorithm (BOA) to enhance the stacking ensemble strategy. By leveraging the strengths of multiple algorithms, BOA dynamically iterates sample weights, thereby amplifying the influence of samples with significant features on the model and improving its capability to capture key pathological characteristics. Following several rounds of parallel experiments, the model proposed in this study demonstrates an accuracy exceeding 80% for normal brain tissue, low-grade gliomas, and high-grade gliomas. This performance reflects a remarkable improvement in accuracy and generalization compared to conventional algorithms. This algorithm holds significant potential for developing portable intraoperative tumor grading instruments, enabling clinicians to formulate more accurate and personalized surgical plans.