FAM-Former: Global Transformer and Feature Aggregation Module for Knee MRI Age Estimation

Age estimation is a significant and active research field in radiological medicine. Recently, researchers have turned their attention to magnetic resonance imaging (MRI) for age estimation, primarily due to concerns regarding ionizing radiation in traditional imaging techniques. In this article, based on the knee MRI dataset, we present a novel end-to-end network that combines convolutional neural network (CNN) and visual transformer network to extract complementary features. In addition, we introduce a feature aggregation module (FAM) that aggregates features from different local knee MRI slices. Our proposed method integrates the feature maps from CNN with the patch embeddings of visual transformer branches, allowing for the acquisition of both local and global information necessary for extracting age-related features. Moreover, we introduce an FAM utilizing the graph attention network (GAT) to achieve interaction between multiple slice features at the feature level. Through extensive experiments conducted on a dataset of 80 knee MRI samples spanning ages 12.0–25.9 years, our method has demonstrated state-of-the-art performance in knee MRI age estimation. Specifically, in a fivefold cross-validation setup, we achieved a mean absolute error (MAE) of 1.52 ± 1.27 years in age regression.

I. Introduction

Age estimation is a highly important topic and a very active research field in both clinical and radiological medicine [1], [2]. In clinical medicine, age estimation plays a crucial role in diagnosing endocrinological disorders, including accelerated or delayed development in adolescents [3]. It also aids in optimizing the timing of pediatric orthopedic surgeries for children approaching puberty, such as procedures for correcting leg length discrepancies or spinal deformities [4]. Artificial age estimation methods, such as those described in [5] and [6], rely on visually observing the ossification of growth plates in bone radiographic images. These methods involve comparing the observed growth patterns with reference data to estimate the age of an individual based on hand or wrist radiographs. Furthermore, referring to the Greulich and Pyle (GP) and tanner-whitehouse 2 (TW2) [7], many studies [8], [9], [10], [11], [12], [13], [14], [15] have summarized the artificial method of knee age prediction based on knee X-ray examinations or magnetic resonance imaging (MRIs). Specifically, the knee images can yield valuable information for three parts: the distal femur, proximal tibia, and proximal fibula [16]. These regions provide essential insights into the ossification stage of the growth plates, which is crucial for age estimation purposes [12]. The process involves categorizing individual slices of a sample into distinct growth stages based on the assessment of skeletal maturation in the knee. Estimating the age of the sample entails aggregating the results obtained from various slices. However, conventional approaches to age assessment in knee age prediction rely on manual analysis performed by expert radiologists, a task that requires extensive professional training. Furthermore, the determination of distinct stages for various image slices in each sample poses a complex challenge, often demanding substantial time and labor investments. When artificially estimating the age of each sample, the age of each slice needs to be separately predicted, which also makes different slices from one sample not correlated at the image level. In addition, even experienced radiologists may encounter challenges such as tedium, time consumption, and subjectivity when diagnosing the age of the knee. These factors can contribute to significant subjective errors, reduced prediction accuracy, and decreased efficiency.