In western countries 90% of Oropharyngeal Cancers are SCC, usually HPV positive [4]: these malignancies tend to cause early lymphatic involvement when compared with HPV negative malignancies.
Sometimes lymphatic metastases are the only sign of malignancy, a well-recognized clinical scenario of “unknown primary tumour” that can account up to 7% of head and neck cancers [5].
Most of these tumors presumably originate from the Oropharynx [6].
Palatine tonsil and tongue base are therefore critical spots for the radiologist in detecting SCC in MRI: morphological changes,
as unilateral tonsillar enlargement,
can be unreliable [2] and sometimes to confirm or exclude the presence of malignancy in this district may be impossible with MRI.
A reliable method to detect within these blind spot is necessary to guide histological sample and to optimize therapy planning.
Bhatia et al showed that mean ADC value can be effectively used for differentiating normal lymphoid tissue from cancer [2].
Ichikawa et al [7] demonstrated that in the oropharynx,
the overall ADCs of lymphomas were significantly lower than those of carcinomas,
probably for the higher number of neoplastic cells,
and,
in contrast,
that the overall ADCs were similar between nasopharyngeal lymphomas and nasopharynx carcinomas.
In a retrospective study,
Choi et al [8] successfully demonstrated that ADC histogram features can detect occult palatine tonsil SCC.
CT histogram features were found to be associated with overall survival in patients with locally advanced head and neck carcinoma who were treated with induction chemotherapy,
regardless of T stage,
N stage,
and other clinical variables [9].
A study by Ravanelli M et Al proved that HPV positive tumours have lower mean ADC values compared with HPV negative ones,
but texture features had a poor discriminatory power for human papillomavirus status [10].
These studies proved that there are significant differences between normal and pathological tissues in terms of ADC value,
that reflects mean value of all the voxel included within a ROI,
but also in texture,
that,
on the contrary,
describe the distribution and relationship of pixel values in the image.
Radiomics is an innovative application in which radiological images are elaborated by software to extract features that quantify a specific aspect of a digital image.
Dimensionality is a critical problem in radiomics, if we consider that many features are highly correlated with each other. Machine Learning (ML) is an effective tool for Radiomics data processing and for combining them to create a predictive model particularly using PCA,
a technique that permits features selection.
The main limit of this study is the small sample size that forced us to use a leave-one-out cross validation: larger dataset can create more reliable models that should be validated on an independent cohort.
Nonetheless,
the present study demonstrated that texture analysis and ML can successfully differentiate SCC from Lymphatic Hyperplasia.
To our knowledge,
this is the first time that ML was used for this purpose.
A ML predictive model could help the radiologist in the hard task of the detection of an unknown primary tumour in head and neck.