Early Discrimination of Periprosthetic Hip Infections Using Neural Networks (SEPTIC-ANNR)

Periprosthetic hip infections are an hot topic in orthopedic surgery, whose incidence is about 1%. The morbidity, mortality and additional costs associated with prolonged hospitalization and further treatments are the main concerns. Periprosthetic infections are generally diagnosed using the CDC (Center for Disease Control and Prevention) criteria (2014). The diagnosis is based on major and minor criteria, including pre-operative and intra-operative parameters. In order to achieve a reliable diagnosis of infection, when fistula are not present, synovial fluid aspiration or tissue samples are required. However, these techniques are expensive and invasive. Moreover, sensitivity is not always so accurate, as shown by some series of revision surgeries performed for presumed aseptic loosening, which turned out to be septic after cultures. Therefore, diagnosis of infection often occurs late and after a long, complex, expensive and not always decisive diagnostic workup, impacting on the timing and success of the treatment.

A practical, rapid, reliable and non-invasive (possibly outpatient) diagnostic procedure for periprosthetic infections would be desirable. It may rely on diagnostic imaging, limiting the collection of liquid or tissues to doubtful cases. Currently, CT and nuclear medicine imaging techniques are not routinely adopted in the diagnosis of infection, due to the modest reliability, costs and exposure to radiant agents.

Recently, neural networks have been introduced: they consist of many simple parallel processors, deeply connected, realizing a computational model. Neural networks mimic brain and its ability to learn. Computational models recognize of visual signals, manage complex situations in real time, classify and manage noise, use associative memory with real-time access to large amounts of data and reconstruct partial or corrupted information. Neural networks have been already used to predict the onset of infections, metastases and treatment failures, integrating clinical and diagnostic imaging data. To date, no studies about neural networks in periprosthetic infection have been conducted. The purpose of this study is to evaluate whether neural networks (cellular neural networks-genetic algorithm), applied to conventional radiographies, are accurate, sensitive and specific for the early-discrimination of a periprosthetic hip infection, already diagnosed with well-recognized methods (CDC 2014).

Specifically, a population of patients, with a complete radiographic history (pre-operative X-rays and a series of other post-operative X-rays), treated for septic or aseptic loosening, is selected.

Both cases are necessary to "instruct" a neural network. The first step consists in identifying a consecutive series of patients with septic or aseptic loosening diagnosis, consulting the hospital database. Thus, patients are categorically divided into septic, or aseptic, loosening. The 2014 CDC criteria are used (as routinely performed in the clinical setting), adding another major and necessary criterion: at least 3 positive intraoperative tissue samples (same micro-organism). In case of aseptic loosening, the case must not meet the CDC 2014 criteria. Thus, the imaging and clinical data of the patients are collected. Having ascertained the diagnosis, the radiographic material is processed (cellular neural networks-genetic algorithm). The proposed procedure processes the radiographic images using the following pipeline and the MatLab software (Mathworks, Natick, US):

• baseline: the first post-implant image is compared to the pre-implant radiographic image;
• progresses are recorded by periodical radiographs using standard and repeatable projections (pelvis X-rays);
• the features are extracted from each image, in the manually segmented area (region of interest - ROI). Three steps take place: 1) image pre-processing, to create uniform frameworks of input data (gray-level images). Color Histogram Equalization; 2) features extracted from neural networks are applied to ROI. Cloning templates using genetic algorithms. The features will be processed by a fully connected layer + SoftMax; 3) features extracted from AutoEconder with fully-connected layer + SoftMax;
• analysis of differential radiographic features (analysis of cellular and convolutive sides) and comparison to the baseline. Post-processing in the following way: 1) fully-connected regression layer and multiclass classifier: will produce the percentage of septic progression risk; 2) fully-connected regression layer and binary classifier: features in the septic / aseptic clusters;
• a final decision tree: fusion of the above-mentioned data, providing a percentage of septic progression risk at the indexed imaging. The aim is to verify whether the neural networks applied to radiographic imaging can accurately, sensitively and specifically recognize a late, chronic periprosthetic hip infection diagnosed according to validated and certain criteria.

Inclusion criteria

• Revisions of primary total hip arthroplasty due to septic and aseptic loosening
• In case of septic loosening, diagnosis of late chronic periprosthetic hip infection
• Complete clinical data
• Complete lab data (pre-revision erythrocyte sedimentation rate and C-reactive protein, at least 5 intraoperative tissue samples).
• Complete radiographic assessment (pre-implant X-ray, a series of post-operative X-rays, pre-revision X-ray)

Exclusion criteria

• Hip re-revisions
• Incomplete or inadequate radiographic assessment
• Inadequate data to diagnose infection according to 2014 CDC criteria and tissue samples