¹⁸F-FDG PET/CT radiomic analysis and artificial intelligence to predict pathological complete response after neoadjuvant chemotherapy in breast cancer patients

Breast Cancer (BC) is still the leading neoplasm in terms of incidence and of cancer-related death in women worldwide [1]. The introduction of neoadjuvant chemotherapy (NAC) has improved the outcomes of BC patients [2, 3]. Literature evidence largely demonstrated that pathological complete response (pCR) after NAC is an early indicator of long-term outcomes [4]. Patients achieving pCR after NAC have significantly longer event-free survival (EFS) and overall survival (OS), in particular in aggressive BC subtypes, such as triple-negative breast cancer (TNBC) [5]. Therefore, the prediction of pCR after NAC through imaging represents a clinically relevant need, as it may improve management and enable treatment personalization. However, available imaging modalities do not guarantee optimal results to solve this task yet. Breast magnetic resonance imaging (MRI) has high accuracy in defining the response to NAC of primary BC [6]. However, breast MRI does not allow whole body staging and de-novo oligometastatic disease could remain concealed. Moreover, false positive (FP) and false negative (FN) findings have been described and the performances are poorer in Luminal subtypes [7‐10]. 2-deoxy-2-[¹⁸F]fluoro-D-glucose ([¹⁸F]FDG) Positron Emission Tomography / Computed Tomography (PET/CT) has been investigated in the last 10 years for imaging BC patients and recently included into guidelines also for the assessment of response to NAC in TNBC and human epidermal growth factor receptor 2 (HER-2) enriched BC [11‐13]. However, also [¹⁸F]FDG PET/CT has limits and the accuracy of conventional PET parameters in the prediction of pCR could be unsatisfactory [14].

Several studies in the last few years focused on the potential information derivable from baseline imaging through the use of radiomic analysis combined with artificial intelligence (AI) in various oncological diseases, including BC [15‐18]. In particular, radiomic analysis of [¹⁸F]FDG PET imaging has been investigated in multiple settings of BC [15]. A large number of papers extracted radiomic features to predict molecular subtypes and histological features of BC at diagnosis, in a modern concept of “digital biopsy” [19]. In other experiences, the radiomic features of axillary lymph nodes were extracted and used to discriminate reactive from metastatic nodes [20]. Weber and colleagues [21] tested radiomics to increase the accuracy of PET imaging to predict the prognosis of BC patients. Moreover, a few papers demonstrated encouraging results by using radiomic analysis extracted by [¹⁸F]FDG PET to predict the response after NAC [22, 23].

Aim of this study is to build machine learning (ML) models able to predict pCR after NAC in BC patients based on conventional and radiomic signatures extracted from baseline [¹⁸F]FDG PET/CT.

The prognostic relevance of pCR after NAC in BC patients has been widely investigated in literature in the last decades [4]. We interrogated PubMed on June 20, 2024 with the following keywords: “pcr and breast cancer and nac”. The results displayed passed from < 5/year in the early 2000s to more than 150/year in 2022 and 2023. The impressive increased number of reports in literature in recent years testifies the great research interest in this parameter. However, pCR is assessed after surgery, thus reducing the therapeutic adjustment only to the initiation of adjuvant treatment in patients with evidence of residual disease. The accurate early prediction of pCR after NAC would be highly relevant to tailor presurgical treatment in BC patients, enabling treatment modulation and potentially improving response rate. For this task, we firmly believe imaging is the main option. Breast MRI and [¹⁸F]FDG PET/CT have been investigated for this purpose [6, 27]. However, post-NAC scans may present false positive or false negative findings [28‐30].

In our previous experience we analyzed volumetric parameters at baseline [¹⁸F]FDG PET/CT and reported that higher values of MTV and TLG of the primary tumor were correlated with residual disease after NAC in Luminal B + HER-2 BC patients [14]. In the current study we confirmed that patients showing higher values of volumetric parameters relative to the primary tumor at baseline [¹⁸F]FDG PET/CT are unlikely to achieve pCR after NAC, regardless of the tumor subtype. In a previous study Evangelista et al. [31] found that higher values of volumetric parameters at baseline [¹⁸F]FDG PET/CT were also associated to shorter PFS in BC patients treated with NAC. Taken together, these results highlight the prognostic relevance of the volumetric extent of the metabolically active burden of disease at [¹⁸F]FDG PET/CT performed before starting NAC in BC patients. However, the value of volumetric parameters is very heterogeneous and tomograph-dependent, thus precluding the possibility to define universal cutoff values that may be used in daily clinical practice.

To overcome the limits of conventional imaging analysis it could be combined with modern radiomic analysis, which enables the collection of hundreds features otherwise hidden to the human eye [32]. Several papers explored the potential of radiomics analysis applied to breast ultrasound and MRI to build AI models able to predict pCR after NAC, with very promising results [33‐35]. Similarly, a few papers tried the same workflow, but on [¹⁸F]FDG PET/CT or PET/MRI [15, 36‐38]. Lee et al. [39] reported suboptimal performances of their models built on clinical-pathological and radiomic features. More encouraging results were reported by Umutlu et al. [22]. However, the authors combined radiomic features of the primary breast tumor, extracted from both PET and MRI images. Recently, Lim et al. [40] developed and externally validated a model built on [¹⁸F]FDG PET/CT radiomic features and significant clinical variables that resulted able to accurately predict pCR after NAC (AUC = 0.729 at external validation). However, once again the model was built using only the radiomic features extracted from the primary breast tumors. In this study, we built a PET Model trained on robust radiomic features extracted from both the primary BC and the reference lymph node lesion, namely PET Model_T + N. The rationale behind this choice is that a pathological uptake at [¹⁸F]FDG PET reflects the metabolic pattern of tumoral cells within the target lesion. Therefore, we hypothesized that the extraction of robust radiomic features also from the most relevant lymph node metastasis would provide useful additional information to the model. Our results look promising as we managed to build ML models able to accurately predict pCR after NAC. Noteworthy, PET Model_T + N outperformed PET Model_T (AUC = 0.83, CA = 0.74, vs AUC = 0.76, CA = 0.70, respectively), confirming our hypothesis that including the RFts extracted from the reference lymph node may improve the performances of the ML model. These preliminary results look promising and should be considered for external validation in upcoming experiences with larger sample size.

As expected, both PET models showed markedly better performances than the CT Model. The CT scan co-registered with PET imaging was a low-dose CT without contrast enhancement (ce), thus suboptimal source of radiomic data in comparison to a diagnostic ceCT. Indeed, focusing on the PET Model_T and the CT Model (both trained only with RFts extracted from the primary breast tumor in the two respective imaging datasets), the first showed better performances probably due to the statistical fluctuation determined by the lack of robust RFts in the CT dataset, producing an overfitting error. Overfitting occurs when a model learns not only the underlying patterns in the training data but also the noise and random fluctuations. As a consequence, the model performs well on the training data but poorly and with low stability on new, unseen data. This is the case of the CT Model, that shows good results in 10-Fold Cross Validation (AUC = 0.89; CA = 0.88) that considerably decrease at validation step (AUC = 0.44; CA = 0.47). Conversely, in PET models the radiomics features represent a large set of quantitative parameters able to discriminate the population and overcoming overfitting at internal validation. However, external validation is needed in future experiences to confirm our promising preliminary results.

Our strategy to include lymph node metastasis only in the radiomic analysis of PET imaging was considered because the metabolic pattern of the tumoral cells within a lesion might be scarcely influenced by the metabolism of the background tissue, both in breast and lymph nodes. Conversely, CT is a morphological imaging and the radiomic features may be influenced by the different physiological density of the tissue considered, unless there is evidence of macroscopic neoplastic subversion of the tissue itself. Therefore, we did not include lymph node CT VOIs into the radiomic analysis.

Considering the clinical Model, the performances resulted suboptimal, as expected considering that clinical-pathological signatures did not result statistically significant to discriminate pCR Vs non-pCR BC in our cohort. Moreover, the relatively small sample size and the unbalanced population did not allow adequate training of the Clinical Model.

A few limitations of our study should be mentioned. Firstly, the retrospective design, the relatively small sample and the lack of external validation represent flaws of the current study that could be overcome in future experiences. Although HER-2 enriched BC are associated to higher response rates than other BC subtypes, HER-2 enriched BC showed a very high rate of pCR after NAC (91%) in our population study, which is higher than that expected according to literature (65%) [41]. However, the ML models presented in this study were trained considering a mixed cohort of BC subtypes. Therefore, the lack of non-responders in the HER-2 enriched group should not have a relevant impact on the reliability of the models presented.

In conclusion, radiomic analysis of baseline [¹⁸F]FDG PET/CT has the potential to contribute to the prediction of pCR after NAC. A validation of our model on external cohort may contribute to personalize the treatment of BC patients unlikely to achieve complete response after therapy.