Intercostal Space Prediction Using Deep Learning In Fully Endoscopic Mitral Valve Surgery

Objective About 10 million people in Europe suffer from mitral valve incompetence. Majority of these entity is mitral valve prolapse in developed countries. Endoscopic mitral valve surgery is a relatively new procedure and preparation in the right intercostal space are crucial for success completion of the procedure. We aimed to explore clinical variables and chest X-rays in order to build most performant model that can predict the right intercostal space for thoracotomy. Methods Overall 234 patients underwent fully endoscopic mitral valve surgery. All patients had preoperative two projection radiography. Intercostal space for right thoracotomy was decided by expert cardiac surgeons taking in consideration the height, weight, chest radiography, anatomical position of skin incision, nipple position and the sex. In order to predict the right intercostal space we have used clinical data and we have collected all radiographies and feed it to deep neural network algorithm. We have spitted the whole data-set into two subsets: training and testing data-sets. We have used clinical data and build an algorithm (Random Forest) in order to have reference model. Results The best-performing classifier was GoogLeNet neural network (now on we will reffera as Deep Learning) and had an AUC of 0.956. Algorithm based on clinical data (Random Forest) had AUC of 0.529 using only chest x-rays. The deep leaning algorithm predicted correctly in all cases the correct intercostal space on the training datasest except two ladies (96.08% ; with sensitivity of 97.06% and specificity 94.12 %, where the Random Forest was capable to predict right intercostal space in 60.78% cases with sensitivity of 93.33% and specificity 14.29 % (only clinical data). Conclusion Artificial intelligence can be helpful to program the minimally invasive cardiac operation, for right intercostal space selection for thoracotomy, especially in non optimal thoraxes (example, obese short ladies). It learned from the standard imaging (thorax x-ray) which is easy, do routinely to every patient.


31
Mitral valve regurgitation (MR) is the second most frequent indication for valve 32 surgery in Europe and it continues to grow. The prevalence of mitral valve prolapse is oscillating between 1 to 2.5 % in developed countries [1]. Usually these are around 60 34 years old patients who expect to have good quality of life and life expectancy is over 35 15 -20 years. Interest in mitral valve repair in fully endoscopic settings is growing [2]. 36 Many groups have reported their results using different types of endoscopic setup [3,4]. 37 However due to large variation of chest dimensions and anatomy there is a discrepancy 38 in choosing the correct thoracotomy intercostal space for given patient. Our 39 experience in mitral valve surgery [5] showed feasibility in the forth intercostal space 40 when using rib spreading and central cannulation (direct vision guided surgery). Many 41 authors choose third intercostal space and some others [4] fourth intercostal space as a 42 default for endoscopic mitral valve surgery (video guided surgery). Intercostal space 43 correct prediction is important of terms of feasibility, short cardio-pulmonary bypass 44 time, short aortic cross clamping (one of most important variables) and overall 45 intervention time. Recently, many approaches for medical imaging analysis have 46 applied deep convulutional networks or 'deep learning' instead of engineering image 47 features [6]. Deep learning is an example of representation learning (without 48 pre-specification of discreiminative features, it learnes for raw data). There are some 49 unwritten rules in different institutions which give some intraoperative tips on how to 50 determine the right intercostal space for thoracotomy; for example, the middle of the 51 thorax measured by both hands (in our institution it is called Farneti's rule, see 52 Figure 1. It is difficult to standardize the approach for given thorax, and it is operator 53 dependent and not reproducible, therefore we attempt here to find a way to predict 54 intercostal space in given thorax.

55
Objective 56 We set a goal to use:    3) All mitral valve replacement cases were considered as a complex mitral valve 75 procedure in endoscopic settings. For complexity see Table 4.   an algorithm that could predict the intercostal space with highest accuracy possible. 135 We have developed the machine learning algorithm using clinical data and some simple 136 radiography measurements; and a deep learning algorithm that could only use the 137 chest radiography for intercostal space prediction.  The size of all the images was readjusted to 2000x2000 with a centered crop operation 158 so as to maintain the DNN input layer with fixed size and partially mask the 159 background that does not provide an useful information for the classification task. 160 Finally, to minimize the computational load for network training, the images were 161 scaled to 512x512 px. Whole data was divided randomly into two parts: train dataset 162 and test dataset. Train dataset was uesed to create the model and validate it. Test  images. For a detailed description of the network architecture see [8].

185
The size of all the images was readjusted to 2000x2000 with a centered crop 186 operation so as to maintain the DNN input layer with fixed size and partially mask 187 the background that does not provide an useful information for the classification task. 188 Finally, to minimize the computational load for network training, the images were 189 scaled to 512x512 px. To evaluate the progress of learning the accuracy was calculated 190 between the predicted and the real "point of incision".  Table 2). Deep learning  Table 3).

207
Using only clinical data it was possible to achieve 60.78% accuracy (range 46.11 -208 74.16 %, see Figure 3 and Table ??). The most important factor for intercostal space 209 prediction was weight, followed by height (see Figure 3 ). There were general 210 agreement on the classification (See Table 3 disagreemend between taking in consideration also gender see Table 3. The 215 best-performing classifier was GoogLeNet neural network (Deep Learning) and had an 216 AUC of 0.9608 (see Figure 4. The deep leaning algorithm predicted correctly in all 217 cases the correct intercostal space on the training datasest except two ladies (96.08% ; 218 with sensitivity of 97.06% and specificity 94.12 % (see Table ??)

220
Mitral valve reconstructive surgery is challenging. It is well known that mitral valve 221 repair can last more then 15 and more years [10,11]. Minimally invasive approach can 222 shorten the hospital stay, less blood product transfusion [5]. In fully endoscopic mitral type of valve surgery [5,12].

229
There is a lot difficulties related to subcutaneous tissue especially when it concerns 230 to obese patients. These are not appreciated when using only weight of the patients or 231 other variable related to it. Big abdomen with severe amount of fat deposits shift the 232 diaphragm in this cases towards lung apex more then it is expected from radiography 233 when anesthetist.

234
Artificial intelligence was started to be applied to medicine and other scientific 235 disciplines [13]. Very recently only the computer become fast enough in order to be 236 fore simple application. Chest radiography is routine procedure for many institutions 237 (if not all for them) and it is easy to do. Using simple simple two projection 238 radiographies with sophisticated algorithm we have reached a very high accuracy. This 239 makes us thing that there are some anatomical substrates hidden in the chest 240 radiography which are possible to extract/make use of it using artificial intelligence.

241
Mentoring in minimally invasive cardiac surgery is a debated issue and it merits a 242 very important attention [14]. Artificial intelligence can be a guide for young surgeons 243 or new programs. We believe that this application of the artificial intelligence could 244 help to find right approach almost for every patient, thus, every newcomer in single 245 mitral valve disease can be addressed to endoscopic program.

247
Artificial intelligence can be helpful to program the minimally invasive, fully 248 endoscopic mitral operation, in order to find the right intercostal space, especially in 249 non optimal thoraxes. It learns from the standard imaging (thorax radiography) which 250 is easy to do routinely and it is cost effective. It is important for mentoring process. 251 Table 1. Schematic representation of the GoogLeNet network used to classify the input images. For a detailed description of the network architecture see [8].