Deep Learning for Pediatric Posterior Fossa Tumor Detection and Classification: A Multi-Institutional Study

Abstract

BACKGROUND AND PURPOSE: Posterior fossa tumors are the most common pediatric brain tumors. MR imaging is key to tumor detection, diagnosis, and therapy guidance. We sought to develop an MR imaging-based deep learning model for posterior fossa tumor detection and tumor pathology classification. MATERIALS AND METHODS:The study cohort comprised 617 children (median age, 92 months; 56% males) from 5 pediatric institutions with posterior fossa tumors: diffuse midline glioma of the pons (n ¼ 122), medulloblastoma (n ¼ 272), pilocytic astrocytoma (n ¼ 135), and ependymoma (n ¼ 88). There were 199 controls. Tumor histology served as ground truth except for diffuse midline glioma of the pons, which was primarily diagnosed by MR imaging. A modified ResNeXt-50-32x4d architecture served as the backbone for a multitask classifier model, using T2-weighted MRIs as input to detect the presence of tumor and predict tumor class. Deep learning model performance was compared against that of 4 radiologists.

Classification of Pediatric Posterior Fossa Tumors Using Convolutional Neural Network and Tabular Data

“…The current work shows an improvement in classification results over a recently published study [16], which was performed in a large study cohort of 617 children. While Quon et al achieved an overall F1 of 0.80, our results show an overall F1 of 0.87 for the validation data and 0.82 for the test datasets.…”

“…Although showing technical feasibility, these studies require a manual phase of tumor segmentation or feature extraction prior to the classification process, which is time consuming and highly dependent on human operator expertise. One study used a deep learning approach in an extremely large multi-institutional cohort of 617 children with PFT, achieving an F1 score of 0.80 [16]. The current study shows improvement over this work, with overall F1 of 0.87 for the validation data and 0.82 for the test datasets.…”

“…While Quon et al achieved an overall F1 of 0.80, our results show an overall F1 of 0.87 for the validation data and 0.82 for the test datasets. This improvement could be explained by (a) the usage of fused architecture, combining CNN and tubular data increase data utilization relative to CNN alone and (b) input data selection; Unlike [16], the best classification results were obtained based on diffusion imaging data, whose values are known to differ between PFT types [3,[7][8][9][10][11][12]. A possible explanation for this is that in our study the diffusion data was used in a form of Trace images, created by averaging the source images.…”

“…A possible explanation for this is that in our study the diffusion data was used in a form of Trace images, created by averaging the source images. In contrast [16] used the calculated apparent diffusion coefficient (ADC) maps which are generated by a mono exponential algorithm with bvalues from 0-1000 s/mm2, varying by institution, and thus might be less robust compared to Trace images. It is important to note that our proposed model was tested on a small dataset, yet outperformed the current state of the art results in this population.…”

“…Several studies have shown the feasibility of PFT classification using conventional machine learning and MRI data. A few studies used histogram textural analysis and visual based features extracted from diffusion and conventional MRI, and patients' clinical features [6,[16][17][18][19][20][21][22][23].…”