Prediction of pituitary adenoma surgical consistency: radiomic data mining and machine learning on T2-weighted MRI

Pituitary adenomas are frequent tumors of the pituitary gland. Although most pituitary macroadenomas have a soft consistency, some are rather fibrous and therefore more challenging to remove by transsphenoidal adenomectomy. Indeed, tumor consistency has been reported as one of the principal determinants of transsphenoidal surgery success rate [1]. For this reason, the ability to preoperatively assess adenoma consistency could improve surgical planning and reduce complication rate and risk of residual tumor presence [2].

Radiomics, consisting of conversion of images into mineable data and subsequent analysis for decision support, has been gaining attention in recent years [3]. In particular, texture analysis is a post-processing technique allowing for quantitative description of pixel gray-level heterogeneity. More recently, texture analysis-derived features have been used in association with data mining and machine learning algorithms, aiding in the interpretation of a large amount of information produced. Machine learning (ML) is the branch of artificial intelligence including algorithms capable of modeling themselves and improving in accuracy by analyzing datasets, without prior explicit programming [4]. It leads to the creation of predictive models that are able, among other tasks, to solve classification problems. The usefulness of the radiomic approach is being assessed in different fields of radiology [5‐9].

Our aim was to assess the accuracy of a ML model trained on radiomic data mined from MRI exams to predict pituitary macroadenoma surgical consistency prior to an endoscopic endonasal procedure.

According to selection’s criteria, 89 patients were included in this study; 51 were males and 38 females, with mean age 52.17 ± 14.64 years (range 16–80). Average lesion size was 25 ± 8 mm (range 8–46 mm). The pituitary lesions were classified as soft in 68 patients and fibrous in the remaining 21. In detail, 19 soft (8 ACTH, 7 GH, 4 PRL) and 6 fibrous (2 ACTH, 2 PRL, 1 GH, 1 TSH) were functioning (25/89, 28% in total). In none of the cases, there was discordance among the neurosurgeons in lesion classification. Patient population clinical data are presented in Table 1.

A total of 1118 texture features were extracted, including first- and higher-order texture features from the original and filtered images. The correlation cluster map of the extracted features is shown in Fig. 3. Their detailed description is available in the online Pyradiomics documentation (https://​pyradiomics.​readthedocs.​io/​en/​latest/​features.​html). After feature stability analysis, 741 were retained for the subsequent steps. Of these, 4 had low variance, while 625 were highly intercorrelated. RFE then identified a 14-feature subset as most accurate (Fig. 4; feature list is available in supplementary material).

The ET model obtained an overall accuracy, in terms of correctly classified lesions, of 86% (± 10%) in the training set cross-validation. The classifier tuned parameters are reported in the supplementary materials. In the test set, the accuracy was of 93% (95%CI = 77–99%), sensitivity of 100%, and specificity of 87%. The AUC is of 0.99 (95%CI = 0.97–1.00) (Fig. 5), equal to the area under the precision-recall curve (0.99), often used in binary ML classifications (fig AUC). The classifier was significantly better (p = 8e⁻⁶) than the no information rate. The confusion matrix and detailed accuracy metrics are shown in Tables 2 and 3.

Preoperative assessment of pituitary macroadenoma consistency is useful for planning surgical approach and reducing residuals and recurrence’s rate. For this reason, several studies have investigated the correlation between preoperative MRI features and tumor hardness. In particular, there are conflicting studies on the value of the relative signal on T2-weighted MRI and the macroadenoma consistency, with some works demonstrating a positive correlation between low signal and hardness [17‐20] and other concluding that relative signal intensity values do not correlate [21‐24]. Indeed besides collagen amount, which mainly correlates with the hardness, other factors such as intratumoral hematoma, amyloid, iron, calcification, or protein-rich fluid may affect the T2 signal intensity [25].

Diffusion-weighted imaging ability to predict tumor consistency also showed divergent results, both indicating a significant correlation [1, 26] and not [21, 27]. Furthermore, the lower spatial resolution and the presence of susceptibility artefacts in the sellar region related to bone and sinus pneumatization limit the use of this technique. Finally, in two studies by Romano et al. and Yamamoto et al., contrast-enhanced MRI showed a strong correlation for tumor consistency [2, 28]. Perfusion imaging parameters have also been investigated as possible biomarkers of pituitary macroadenoma consistency, but no added value was found compared with precontrast T1-weighted images [29]. A more interesting advanced technique in this setting is represented by MR elastography. Pituitary adenoma stiffness was found to correlate with their consistency and, if it became widely available, could offer additional data to mine with a radiomic approach [30, 31].

Regarding texture analysis, there are only three studies exploring this issue, to the best of our knowledge. In the first, Rui et al. explored the value of MRI texture analysis in assessing pituitary macroadenoma consistency, obtaining good accuracy values [32]. However, this study was conducted using contrast-enhanced 3D-SPACE images and without a ML approach. Fan and colleagues explored this issue in acromegalic patients using ML for radiomic feature selection prior to building a nomogram obtaining an AUC of 0.81 [33]. Zeynalova et al. also performed an analysis on ML preoperative evaluation of pituitary macroadenoma consistency [34]. Their study presented some similarities with our own. They also used Pyradiomics for feature extraction from bidimensional ROIs, although they utilized different sigma settings (2, 4, and 6 mm) for the LoG filter and obtained a total of 162 parameters. The lower number of features is probably due to their exclusive focus on first-order histogram-derived ones. These are more reproducible but convey less information on tissue texture compared with higher-order parameters. As their in-plane resolution was higher (0.5 × 0.8 mm), they were able to use a 1 × 1 mm resampling size compared with our 2 × 2 mm. It is interesting to note the use of a very narrow bin width value of 0.06, as the number of bins should not exceed 128, following the developer recommendations. They also performed a feature robustness assessment with our same ICC threshold, while their intercorrelation threshold was lower (0.7 vs 0.8). After data dimensionality reduction, they identified 6 informative features using the Weka data mining platform and a wrapper-based selector. In our study, the entire analysis was conducted using the scikit-learn Python package. Some other major differences are represented by the use of cross-validation, without further assessment on a separate test set. Their reported accuracy is 72.5%, with an AUC of 0.71. Therefore, our algorithm presents a clearly superior performance. This could be in part explained by their use of a multilayer perceptron neural network, which may not be the best suited algorithm for a small dataset obtained from 55 patients. Finally, Zenyalova and colleagues also used collagen amount within the tumor on histopathological examination for their reference standard. As consistency information is mainly useful for surgical strategy planning, we believe that intraoperative consistency assessment represents a more practical and useful reference standard as the final recipient of the information should be a neurosurgeon. As the two neurosurgeons involved in our study never had disagreements, we also found this assessment to be reproducible.

By analyzing our confusion matrix, it can be seen that the mistakes made by the classifier were 2 cases of soft lesions identified as fibrous. Given the clinical setting of our investigation, this kind of error is somewhat more acceptable than a false negative, as it would be more auspicable to sometimes overestimate the difficulty of a surgery rather than the opposite.

In our study, we chose to employ an ET ML algorithm. This belongs to the decision tree ensemble methods, in particular constituted by a large number of highly randomized decision trees which are fitted on data subsamples. Each of these outputs a prediction, and a majority vote determines the final outcome. Ensemble learning is based on the assumption that a decision by committee made by a large number of weak classifiers will perform better than a single algorithm. A sufficient diversification of the random trees included in the ET is guaranteed by random sampling, with replacement, of patients from the training dataset (bootstrap aggregation or bagging) and of their available features (n = 3 in our case). This in turn ensures low correlation of each tree, improving the ET’s overall performance [16]. As the dataset lesion classes were imbalanced, SMOTE was employed. This is a known solution to address this issue and has demonstrated its value in the setting of medical imaging radiomic ML analysis [15, 35‐37].

We have chosen a handcrafted radiomics approach rather than a fully automated deep learning one as this gave us better control on the initial data analysis and following ML model construction. Both approaches have been object of discussion in current literature as they possess peculiar merits and limitations. It is our belief that a handcrafted analysis is more appropriate for relatively smaller datasets as it allows greater involvement of radiologists and better understanding of the whole pipeline. Only when extremely large datasets will become available in medical imaging, as in other fields, the less time-consuming completely neural network-based approach will be a practical necessity. Until then, the value of greater involvement of the radiologist and finer quality control of patient or lesion data outweigh the larger amount of time needed to extract medical imaging radiomics. Furthermore, medicine and especially treatments are evolving in the direction of precision, patient-tailored therapies. Contrary to the current desire in radiology to aggregate as many patients as possible to train ML algorithms, this determines a need to work with ever smaller patient subgroups within each pathological entity. Therefore, a future with space for both deep learning software to apply on large populations and engineered approaches for more specific tasks can be envisioned.

Our study has some limitations which have to be acknowledged. As is often the case for ML, future studies on larger populations are necessary to confirm and possibly expand our results. The need for oversampling given the unbalanced nature of the classes further highlights this necessity but was expected given epidemiological data. Only T2-weighted images were used, without investigating the added value of other sequences. However, obtaining valuable data without contrast agent administration and streamlining the pipeline to incorporate a single MRI sequence could also represent an added value. Furthermore, considering previous works, T2-weighted MRI alone proved effective to provide data concerning proliferative index [38], secretory activity [39], and response prediction to somatostatin analogues in patients with acromegaly and GH secreting pituitary macroadenoma [40, 41].

The ML model trained on radiomic data extracted from T2-weighted MRI demonstrated a high accuracy in the classification of soft and fibrous pituitary macroadenomas. Therefore, this tool could prove valuable in the pre-surgical planning of these patients if further developed and validated on larger datasets.