Editorial for “Pretreatment Multiparametric MRI-Based Radiomics Analysis for the Diagnosis of Breast Phyllodes Tumors”

Phyllodes tumors (PTs) are rare tumors that comprise 0.3%–1% of all primary breast tumors. The name is of Greek origin meaning “leaf,” referring to the leaf-like morphology of the tumor. Per recommendations by the World Health Organization (WHO),¹ PTs are classified into benign, borderline, and malignant based on a constellation of histological parameters. PTs are treated by surgical excision. Pathological tumor grade determines the choice of surgical method and the extent of surgical excision. While there is no consensus on the definition of an appropriate surgical margin, margin status is a strong predictor of local recurrence. Risk of recurrence and metastasis also increases with increasing tumor grade. Even though the number of deaths represented a small percentage of the total population, all of the PT deaths (7) reported by Tan et al² were preceded by a malignant primary diagnosis. Benign tumors do not metastasize and are associated with a lower risk of local recurrence. However, the determination of the aggressiveness of PTs is challenging due to overlapping histologic features between each grade, the subjective nature of assessed histologic parameters, and the need to assess and integrate multiple parameters into an accurate diagnosis. Achieving an accurate diagnosis is important to help predict recurrence potential and guide clinical management.

Improving the diagnostic accuracy of PTs is an active area of research. Ultrasound and mammography are first-line, standard modalities for screening and diagnosing PTs, but magnetic resonance imaging (MRI) has been increasingly performed based on its superior resolution and improved visualization, providing added value in PTs staging. MR findings of PTs grade have been reported,³ but the diagnosis requires extensive expertise and imaging impressions are based on subjective assessments. There have been an increased number of reports applying medical image processing radiomics analysis to MR images for the characterization of PTs.^4–7 These studies demonstrate the potential of radiomics analysis for differentiating the pathological grades of PTs and their aggressiveness. While earlier studies had limited datasets (probably due to the rarity of these tumors) and explored a few number of radiomics features, recent studies had larger patient populations and were able to extract a higher number of radiomics features.^7,8

In this issue of JMRI, Ma and colleagues⁹ demonstrated an approach for improving the accuracy of pretreatment differential diagnosis of PTs through the incorporation of MR findings, radiomics extracted from T1W, T2W, and DCE-T1W images, and clinical data for each subject. Ma et al reported the result of a retrospective, single-center study in China, enrolling 216 patients with clear grading based on pathologically evaluation of PTs. The best performing model at differentiating between benign PTs (BPTs) and borderline/malignant PTs (BMPTs) is based on 12 parameters, a combination of radiomics features extracted from multiparametric MR images and clinical features including age, individual history of breast tumor, individual history of other tumors, and family history of breast tumor. The data presented in table 3 show a marginal improvement in area under the curve (AUC) in the validation dataset compared to other models evaluated. The improvement seen with the combined clinical features and multiparametric MRI radiomics agree with the recent trend to combine radiomics and clinical data for disease/tissue characterization. Considering that this is a single-center trial, multicenter studies need to be performed to validate the approach.

The study provides valuable insight in the MRI characteristics for differentiating BPTs from BMPTs in a large population cohort. In line with previous reports, tumor size and heterogeneous signal on T2W are observed more frequently in BMPTs than in BPTs. This agrees with the general observation that malignant PTs have a marked degree of hypercellular stromal overgrowth, proliferation in the stroma, and an invasive border. The authors listed variability in MRI scanners (vendor and field strength) as part of study limitation. Indeed, scanner hardware and software variation introduced major confounding factors and may affect the reproducibility of radiomics feature. However, in radiomics studies, pooling images acquired using different scanners is needed to increase the size of the cohort. With an appropriate image analysis protocol, inclusion of a wide variety of images may assist in the generalization of the radiomic models. Like most of the recent studies, regions of interest (ROIs) in this study were manually drawn. Incorporating automatic segmentation in image analysis pipeline would lessen the subjectivity and may improve clinical translation potential.

In summary, the study by Ma et al adds to the sparse body of knowledge on the link between MRI findings, multiparametric MR radiomics features, clinical characteristics, and pathological outcomes of PTs. While the model is still dependent on the subjectivity of the radiologist, it shows potential in the development of an accurate diagnosis model for guiding the clinical management of PTs.