In endometrial cancer (EC), one of the most important factors for determining how aggressive the disease is and how it should be treated is the depth of myometrial invasion (DMI) - how far the tumor has grown into the muscular wall of the uterus (the myometrium). Under international staging guidelines, invasion is classified into three categories: no or only focal invasion; less than 50% of the myometrial thickness; or 50% or more (deep invasion).
This measurement matters because it directly shapes treatment decisions. In patients with no or shallow invasion, less extensive surgery is appropriate, and routine removal of lymph nodes may not be needed. In patients with deep invasion, the risk of cancer spread to lymph nodes increases, and more extensive surgery (including systematic lymphadenectomy) may be recommended. Deep myometrial invasion also triggers additional post-surgical treatments such as radiation therapy.
Currently, the two main imaging tools used to assess myometrial invasion before surgery are magnetic resonance imaging (MRI) and ultrasound (US). Both have similar overall accuracy, but ultrasound is far more accessible, less expensive, faster, and causes less discomfort to patients. The major limitation of ultrasound, however, is that it requires manual interpretation and its accuracy depends heavily on the experience of the individual operator performing the scan - making results variable between clinicians.
Radiomics offers a solution: by extracting large numbers of quantitative features from medical images using algorithms rather than human eyes, it can capture subtle patterns in image texture, shape, and structure that are invisible even to experienced operators - and do so consistently and objectively every time.
The researchers collected ultrasound images from 77 consecutive patients with confirmed endometrial carcinoma treated at three Italian cancer centers between 2021 and 2023. Transvaginal ultrasound scans were performed by experienced operators (level II or III certification), and the images were then analyzed using a computational pipeline. The surgical pathology findings from hysterectomy specimens served as the definitive ground truth for training and evaluation.
Image preprocessing was an important first step: since ultrasound images often contain measurement markers (lines, arrows, text) added by radiologists during the scan, an inpainting algorithm was used to automatically remove these markers and reconstruct the underlying image texture, preventing the AI from inadvertently learning from these artifacts.
Features were extracted using Inception-V3, a deep convolutional neural network originally trained on millions of natural images (not medical images). Using this pre-trained network as a feature extractor - a technique called transfer learning - allowed the researchers to work with a small clinical dataset without needing to train a deep network from scratch. From each image, 235,200 features were extracted from an intermediate layer of the network that captures low-level structural details such as edges, curves, and texture gradients.
To reduce these 235,200 features to a manageable and relevant subset, a two-step feature selection process was applied: first, filtering by statistical association with the outcome (Kruskal-Wallis test), then applying the Maximum Relevance - Minimum Redundancy (MRMR) algorithm to select features that best predict the outcome while minimizing overlap with each other. The final selected features were used to train Support Vector Machine (SVM) classifiers - both a binary model (shallow vs. deep invasion) and a three-class cascade model (focal/none, less than 50%, and 50% or more).
The binary classification model (predicting shallow vs. deep invasion) achieved a mean AUC of 90.76% across 100 rounds of 10-fold cross-validation. Overall accuracy was 86%, with 85% sensitivity and 91.3% specificity. In comparison, experienced human operators performing the same binary task achieved an overall accuracy of 80.5%, with precision of 73.5% - considerably lower on the key metric of how often flagged cases were truly positive.
When operators were given access to the radiomic model's predictions as a decision-support tool, their precision improved significantly to 86.7% (compared to 73.5% without support), a statistically significant improvement (p less than 0.05). This demonstrates the potential of AI-assisted reading to meaningfully upgrade human diagnostic accuracy in clinical practice.
The three-class (multiclass) model achieved an overall accuracy of 84.93% and a Macro F1-score of 84.5%. Performance varied by class: recall was 93.3% for focal/no invasion, 71.9% for less than 50% invasion (the intermediate category, which is naturally the hardest to classify), and 90% for deep invasion (50%+). Unaided operators achieved only 74% accuracy on the three-class task, with a notably low recall of 56.3% for the intermediate class - the exact subgroup where misclassification most affects treatment decisions.
With AI assistance, operator accuracy on the three-class task rose from 74% to 82.8%, and recall for the critical intermediate class rose from 56.3% to 71.9% - both statistically significant improvements. This highlights the greatest clinical value of the tool: helping clinicians correctly identify the intermediate invasion group, where decisions about lymph node staging are most ambiguous.
Correctly distinguishing deep from shallow myometrial invasion before surgery has direct implications for what type of operation a patient undergoes. Guidelines recommend that lymph node removal (lymphadenectomy) be performed when deep invasion is present or suspected, because cancer is more likely to have spread. In patients with clearly shallow invasion, lymphadenectomy can be avoided, sparing patients from an invasive procedure with its own risks and side effects.
The intermediate class - invasion present but less than 50% - is particularly important to identify accurately because it is where clinical uncertainty is highest. A tool that consistently and objectively classifies these cases reduces reliance on individual operator skill and helps standardize care across different hospitals and healthcare settings.
Beyond lymph node staging, accurate preoperative invasion assessment influences decisions about postoperative radiation therapy, and it plays a role in identifying young patients who might be candidates for fertility-sparing treatment. For women who wish to preserve their ability to have children, confirming the absence of myometrial invasion on imaging is one requirement for consideration of non-surgical management options.
This study demonstrates that a radiomic model based on deep learning feature extraction from ultrasound images can achieve high accuracy in predicting myometrial invasion depth in endometrial cancer - matching or exceeding the performance of experienced ultrasound operators, and meaningfully improving operator performance when used as a decision support tool.
The approach is particularly valuable because it works with ultrasound rather than MRI, making it far more accessible globally. Ultrasound equipment is available in virtually all hospitals and outpatient clinics, and transvaginal ultrasound is routinely performed as part of initial EC workup. Adding AI-based radiomic analysis to an existing clinical workflow requires no additional imaging procedures.
A key strength of the radiomic approach is objectivity and standardization: unlike a human reading the same scan differently on different days, the algorithm produces the same analysis from the same image every time. This addresses one of the fundamental weaknesses of ultrasound in cancer staging - its dependence on operator expertise and experience.
The main limitation of this study is the small sample size (77 patients), which necessitated careful cross-validation but limits generalizability. Future work should validate the model on larger, multi-center datasets and across different ultrasound scanners and patient populations before clinical adoption. Integration into routine clinical workflows as a real-time decision-support tool during scanning represents the next step.