Skip to main content
  • Rapid Intraoperative Differentiation of Pseudoprogression and Glioma Recurrence using Stimulated Raman Histology

    Final Number:

    Todd Charles Hollon MD; Balaji Pandian; Sudharsan Srinivasan; Esteban Urias; Daniel A. Orringer MD

    Study Design:

    Subject Category:

    Meeting: Congress of Neurological Surgeons 2018 Annual Meeting

    Introduction: Accurate intraoperative differentiation of tumor progression and pseudoprogression is essential in the surgical management of recurrent gliomas. Reliably differentiating recurrent tumor versus pseudoprogression can be challenging and time-intensive with conventional intraoperative histology. Stimulated Raman histology uses the intrinsic biochemical properties of fresh, unprocessed surgical specimens to provide rapid label-free virtual histologic images. Here, we report a deep convolutional neural network (CNN), which differentiates pseudoprogression and recurrent glioma in fresh surgical specimens imaged with SRH in a fraction of the time required for conventional histologic techniques.

    Methods: For CNN training, 1.1 million 400x400µm SRH fields of view (FOV) from 405 patients was used to build a robust multiclass CNN. Our deep neural network was developed from the GoogleNet InceptionV3 CNN architecture, which includes 24 million trainable parameters. For model testing, 36 previously treated glioma patients were used to determine classification accuracy. The model was evaluated on its ability to diagnose 1) normal brain 2) pseudoprogression/gliosis/treatment effect, 3) low-grade glioma (LGG) recurrence, 4) malignant glioma recurrence, and 5) overall diagnostic accuracy.

    Results: Time to acquire a SRH image was less than 1 minute. The trained CNN achieved an overall diagnostic accuracy of 97.2%. When evaluating each imaged specimen, the model was able to diagnose normal brain (10 patients), pseudoprogression (6 patients), LGG (8 patients), and malignant glioma (12 patients) with 94.1%, 96.3%, 95.0%, 88.6% accuracy, respectively. The only classification error made by the CNN was an instance of dense gliotic brain tissue due to treatment effect being incorrectly classified as recurrent LGG.

    Conclusions: Our study demonstrates the feasibility of applying deep learning for intraoperative diagnosis of pseudoprogression and recurrent glioma in SRH imaged tissues. In the future, convolutional neural networks may ultimately be used to guide decision-making in the surgical care of recurrent gliomas, independent of conventional neuropathology resources.

    Patient Care: Our research will improve patient care by increasing the accuracy and reliability of differentiating recurrent brain tumors versus pseudoprogression using a combination of rapid SRH imaging and deep convolutional neural networks. Our research utilizes the most recent advances in optical imaging and artificial intelligence to improve the intraoperative management of brain tumors.

    Learning Objectives: By the conclusion of this session, participants should be able to: 1) Describe the importance of differentiating tumor progression versus pseudoprogression using SRH images analyzed using deep convolutional neural networks 2) Discuss, in small groups, how modern artificial intelligence techniques can improve intraoperative diagnosis and detection of tumor infiltration.


We use cookies to improve the performance of our site, to analyze the traffic to our site, and to personalize your experience of the site. You can control cookies through your browser settings. Please find more information on the cookies used on our site. Privacy Policy