Medulloblastoma tumor classification using deep transfer learning with multi-scale EfficientNets

Abstract

Medulloblastoma (MB) is the most common malignant brain tumor in childhood. The diagnosis is generally based on the microscopic evaluation of histopathological tissue slides. However, visual-only assessment of histopathological patterns is a tedious and time-consuming task and is also affected by observer variability. Hence, automated MB tumor classification could assist pathologists by promoting consistency and robust quantification. Recently, convolutional neural networks (CNNs) have been proposed for this task, while transfer learning has shown promising results. In this work, we propose an end-to-end MB tumor classification and explore transfer learning with various input sizes and matching network dimensions. We focus on differentiating between the histological subtypes classic and desmoplastic/nodular. For this purpose, we systematically evaluate recently proposed Effi-cientNets, which uniformly scale all dimensions of a CNN. Using a data set with 161 cases, we demonstrate that pre-trained EfficientNets with larger input resolutions lead to significant performance improvements compared to commonly used pre-trained CNN architectures. Also, we highlight the importance of transfer learning, when using such large architectures. Overall, our best performing method achieves an F1-Score of 80.1%.

Multi-Scale Input Strategies for Medulloblastoma Tumor Classification using Deep Transfer Learning

“…We consider MB subtype classification using pre-trained Effi-cientNets and study the impact of input patches with different scales and global context for this task. Our results in Table 1 demonstrate that using the previous approach [3] with larger tiles downsampled to the network input resolution only lead to minor performance improvements for a tile size of 4000 × 4000 px. However, for extremely large tiles (8000 × 8000 px) performance is even reduced.…”

“…We follow the concept of a previous study on MB classification [3] and consider pre-trained EfficientNets [18]. The key advantage of EfficientNets is the compound scaling method, which uniformly scales network width, depth, and input resolution starting with the baseline EfficientNet-B0.…”

“…The key advantage of EfficientNets is the compound scaling method, which uniformly scales network width, depth, and input resolution starting with the baseline EfficientNet-B0. Considering the findings of previous works [3], we focus on EfficientNet-B0 (E#Net-B0) with an input resolution of 224 × 224 and EfficientNet-B5 (E#Net-B5) with an input resolution of 456 × 456. Note, we use architectures pre-trained on ImageNet.…”

“…Our general classification pipeline is shown in Figure 1. Given the sets with different tile sizes, we first follow the previous approach [3] and simply downsample an entire image tile to the corresponding input resolution (h i × w i ) of a CNN. This leads to extreme downsampling of the larger tiles, especially for 8000 × 8000.…”

“…To counter this problem, transfer learning is typically used in digital pathology [9,14]. To mitigate the problem of scant training examples in MB classification, a recent study has compared various benchmark CNN architectures together with transfer learning [3]. The study demonstrates that Efficinetnets [18] with a larger input resolution outperform classical CNN architectures with smaller input resolutions.…”