Acute leukemia (AL) is a cancer of the blood characterized by the rapid accumulation of immature blood cells. While many patients achieve complete remission after initial chemotherapy, relapse remains a major cause of treatment failure and death.
When leukemia returns, it can do so within the bone marrow (intramedullary), in other tissues outside the marrow (extramedullary), or both simultaneously. Monitoring for relapse therefore requires ongoing surveillance of both locations.
Bone marrow biopsy (BMB) is the gold-standard test for detecting relapse inside the marrow, but it is invasive, painful, and only samples a small fraction of the entire bone marrow. A non-invasive, whole-body imaging method would be far preferable.
18F-FDG PET/CT is a nuclear medicine scan that combines positron emission tomography (PET), which detects metabolically active cells, with computed tomography (CT), which provides anatomical detail. Cancer cells consume glucose at high rates, making them visible on FDG-PET scans. This technology has been shown to detect extramedullary lesions that routine tests miss.
When a radiologist looks at a PET scan, they assess bone marrow involvement (BMI) by judging whether the sugar uptake in the marrow is higher than expected. The challenge is that this uptake can appear as focal (localized hot spots), normal (uptake equal to or lower than the liver), or diffuse (uniformly elevated throughout the marrow).
Diffuse uptake is both the most common pattern in leukemia patients and the most ambiguous, because it can look identical whether caused by leukemic cells or benign conditions such as infection, anemia, or recovery from chemotherapy. This ambiguity places a heavy burden on the physician's experience and judgment.
Radiomics is a field that extracts hundreds or thousands of quantitative features from medical images - textures, shapes, intensity patterns - that are invisible to the human eye. These features can potentially capture disease signatures that experienced radiologists cannot detect visually. The researchers hypothesized that machine learning applied to radiomic features could overcome the limitations of visual diagnosis.
The study retrospectively included 41 patients with suspected relapsed acute leukemia (both AML - acute myeloid leukemia and ALL - acute lymphoblastic leukemia) who underwent 18F-FDG PET/CT scanning at Peking University People's Hospital between 2012 and 2019.
All patients also underwent bone marrow biopsy within one week of their scan, and the biopsy result served as the definitive ground truth (gold standard) for whether bone marrow involvement was truly present or absent. Of the 41 patients, 18 were biopsy-positive and 23 were biopsy-negative.
The dataset was split into a training group (35 patients from 2012-2018) used to build the model, and an independent validation group (6 patients from 2018-2019) used to test whether the model worked on completely new cases it had never seen before. Three experienced nuclear medicine physicians independently performed visual analysis of all scans as a comparison baseline.
The researchers manually outlined the volumes of interest (VOIs) - the spine and pelvis - on each patient's PET and CT images. From these regions, a software package automatically extracted 1,826 quantitative features: 781 from PET and 1,045 from CT. These features captured intensity distributions, texture patterns, shapes, and spatial characteristics at multiple frequency scales using wavelet and Laplacian-of-Gaussian image filters.
Starting with 1,826 features and only 35 training patients creates a severe risk of overfitting - where a model memorizes the training data but fails on new patients. To address this, the team used a multi-stage feature selection pipeline: first, consensus clustering identified groups of highly correlated features; next, a random forest algorithm ranked each feature by its predictive importance; then, only features with area under the ROC curve (AUC) above 0.70 were retained; finally, recursive feature elimination removed redundant features.
This rigorous process whittled 1,826 features down to just 3 key features, all derived from wavelet-transformed images: two PET texture features (run entropy and kurtosis) and one CT texture feature (short run high gray level emphasis). These captured subtle patterns of tissue heterogeneity that the eye cannot detect.
The final prediction model used a Random Forest algorithm - an ensemble method that builds many decision trees on random subsets of data and averages their predictions. Random forests are particularly well-suited for small datasets with many features because they are robust against overfitting and handle high-dimensional data efficiently.
Model performance was assessed using two complementary approaches. Ten-fold cross-validation divided the 35-patient training set into 10 parts, training on 9 and testing on 1 repeatedly, which gives a reliable estimate of internal performance. Independent validation then applied the fully trained model to the 6 patients it had never seen, providing the most honest test of real-world generalizability.
Performance was measured using sensitivity (the ability to correctly identify true positive cases), specificity (the ability to correctly rule out negative cases), overall accuracy, and the area under the ROC curve (AUC), a single number summarizing the model's ability to discriminate between positive and negative cases across all possible thresholds.
Visual analysis by experienced physicians achieved an overall accuracy of 68.6% with a sensitivity of 62.5% and specificity of 73.7% (AUC = 0.681). These numbers reveal that even expert radiologists correctly identify bone marrow involvement only about two-thirds of the time when relying on visual inspection alone.
The machine learning model achieved 88.6% accuracy in cross-validation, with sensitivity of 87.5% and specificity of 89.5% (AUC = 0.885). This was statistically significantly better than visual analysis (P=0.041 for accuracy; P=0.046 for AUC), meaning the improvement was very unlikely to be due to chance.
The improvement was most dramatic for the clinically challenging diffuse uptake cases - the most common and ambiguous pattern. Physicians correctly classified only 58.3% of diffuse uptake patients, while the machine learning model achieved 83.3% accuracy on the same cases, correctly rescuing 9 out of 10 cases that physicians had misdiagnosed.
On the completely independent validation group of 6 patients, the model achieved 83.3% accuracy, demonstrating that its performance was not simply memorization of the training data but a genuinely generalizable capability.
The three selected features all came from wavelet-transformed images - mathematical decompositions that separate image information into different frequency bands, revealing textural patterns at multiple scales. This suggests the key diagnostic signal lies in mid-frequency texture patterns rather than simple overall brightness.
Run Entropy (PET) measures the randomness in the distribution of gray-level runs - sequences of pixels with the same intensity. Higher entropy means more chaotic, heterogeneous texture patterns. BMB-positive patients showed higher run entropy on average, suggesting malignant marrow involvement creates more disordered metabolic activity patterns.
Kurtosis (PET) measures the peakedness of the intensity distribution - how concentrated pixel values are around the mean. BMB-negative patients showed higher kurtosis, meaning their marrow uptake was more uniformly concentrated around a typical value, as expected in benign conditions. Leukemic involvement disrupts this regularity.
Short Run High Gray Level Emphasis (CT) captures the prevalence of short runs of high-intensity pixels in the CT image. This feature was lower in BMB-positive patients, suggesting that the bone structure appears more heterogeneous at a fine scale when marrow is involved. Notably, conventional metrics like SUVmax and metabolic tumor volume performed far worse than these wavelet texture features.
A common criticism of machine learning in medicine is that models are black boxes - they make predictions without explaining why, which makes clinicians reluctant to trust them. The researchers addressed this by applying the LIME (Local Interpretable Model-agnostic Explanations) framework, which approximates the complex model as a simple linear combination of weighted features for each individual patient.
This allowed the researchers to show, for any given patient, which features pushed the prediction toward positive or negative BMI and by how much. For example, in one illustrative case, the run entropy feature pushed strongly toward a BMI-negative prediction while kurtosis pushed toward positive, and the overall balance determined the final output.
This kind of per-patient interpretability is clinically valuable because it allows physicians to understand the reasoning behind a specific prediction, identify edge cases where the model may be less reliable, and build the trust necessary for eventual clinical adoption.
Bone marrow biopsy, while definitive, is an invasive procedure that samples only a tiny portion of the entire bone marrow. Patients with suspected relapse often undergo multiple biopsies during follow-up, creating significant discomfort and procedure-related risks. A reliable imaging-based test could reduce the need for repeat biopsies.
The radiomic model correctly predicted 10 out of 11 cases that visual analysis had failed to diagnose. This suggests it could serve as a complementary non-invasive test - not replacing biopsy entirely, but helping physicians decide when biopsy is truly necessary and reducing unnecessary invasive procedures when imaging evidence is strongly negative.
The researchers plan to extend this work by automating bone segmentation (currently done manually, which is time-consuming), combining it with the prediction model into a fully automated pipeline, and validating it across multiple hospitals. Multi-center validation is essential before any clinical tool can be broadly adopted, as it tests whether the model works across different scanner types, patient populations, and clinical practices.
This study demonstrated for the first time that radiomics combined with machine learning can significantly improve the diagnosis of bone marrow involvement in suspected relapsed acute leukemia patients, achieving nearly 89% accuracy compared to 69% for experienced physician visual reading.
The approach is particularly valuable for diffuse uptake patterns, which are the most common and most ambiguous presentation in leukemia, and where visual diagnosis is least reliable. By quantifying subtle texture features invisible to the eye, machine learning overcomes the inherent subjectivity of visual interpretation.
Future steps include building fully automated bone segmentation tools, deploying the system in collaborative hospitals for multi-center validation, and ultimately standardizing imaging biomarkers for bone marrow assessment. If validated at scale, this technology could meaningfully change how leukemia relapse is monitored, making it faster, less painful, and more accurate for patients worldwide.