Robust deep learning classification of adamantinomatous craniopharyngioma from limited preoperative radiographic images

Abstract

Deep learning (DL) is a widely applied mathematical modeling technique. Classically, DL models utilize large volumes of training data, which are not available in many healthcare contexts. For patients with brain tumors, non-invasive diagnosis would represent a substantial clinical advance, potentially sparing patients from the risks associated with surgical intervention on the brain. Such an approach will depend upon highly accurate models built using the limited datasets that are available. Herein, we present a novel genetic algorithm (GA) that identifies optimal architecture parameters using feature embeddings from state-of-the-art image classification networks to identify the pediatric brain tumor, adamantinomatous craniopharyngioma (ACP). We optimized classification models for preoperative Computed Tomography (CT), Magnetic Resonance Imaging (MRI), and combined CT and MRI datasets with demonstrated test accuracies of 85.3%, 83.3%, and 87.8%, respectively. Notably, our GA improved baseline model performance by up to 38%. This work advances DL and its applications within healthcare by identifying optimized networks in small-scale data contexts. The proposed system is easily implementable and scalable for non-invasive computer-aided diagnosis, even for uncommon diseases.

Application of Artificial Intelligence in Diagnosis of Craniopharyngioma

“…(5) Prince et al ( 30 ) studied a series of optimization methods to address small data sets and subsequently adopted deep learning algorithm to identify the pediatric adamantinomatous craniopharyngioma. Transfer learning was an effective technique to deal with overfitting of small data sets, which was verified in the study.…”

“…For example, Carver et al ( 82 ) used GAN network to generate high-quality images. Price et al ( 30 ) employed two data augment techniques, one was a random process that used probability thresholds for sample transformations, and the other was a transformation adversarial network for data augmentation (TANDA). The relative simplicity of the image led to the advantage of random augment over TANDA.…”

“…Mehmet et al ( 81 ) adopted a CNN network to automatically detect high-grade glioma and achieved good performance. Prince et al ( 30 ) adopted a CNN network with parameter optimizer to realize the non-invasive diagnosis of adamantinomatous craniopharyngioma. Wu et al ( 77 ) employed three CNN network models (AlexNet, ResNet, Inception-V3) to classify glioma and encephalitis.…”

Molecular Biology in Treatment Decision Processes—Neuro-Oncology Edition

“…Data collected is represented by direct data entry fields (manual or automatic) in the treatment planning system (e.g., tumor site, treatment intent, ready to treat dates, patient setup and desired technique employed in treatment planning), as well as tumor and normal tissue volumes delineated by the clinician and RT dose delivered using multiple DVH (Dose Volume Histogram) (Figure 1). These large-scale datasets can be employed for administrative purposes, such as capturing the number of patients on treatment that share a common histology or planning technique, but are also most relevant to computational approaches that involve artificial intelligence (AI) [3,4,[9][10][11][12][13], machine learning (ML) [2,[4][5][6][7]9,[13][14][15][16][17][18][19][20][21][22], deep learning (DL) [2,3,11,[23][24][25][26][27], ground truth [7,13] and radiogenomics [2,13,14,[28][29][30][31][32][33][34] (See Table…”

“…Due to the rare nature of these tumors, large-scale real-world data is mostly lacking, and randomised data are almost impossible to acquire making computational approaches especially relevant. Most studies have focused on diagnosis in the context of CNS tumors where tissue acquisition is challenging or impossible such as pediatric posterior fossa tumors [19], rare histologies or histologies that present difficult diagnostic interpretation (ependymoma, pilocytic astrocytoma, medulloblastoma, craniopharyngioma) [10,20,27,78,79] and very few studies examined computational avenues to optimise RT [80], Zhu [81]. In meningioma attempts have focused on diagnosis and grading especially in the context of radiomics and surgical resection [21] and linked analysis to extent of tumor and brain or bone invasion [82][83][84][85].…”