Quick and reliable diagnosis of stomach cancer by artificial neural network

  • Authors:

  • Saeid Afshar;Fahime Abdolrahmani;Fereshte Vakili Tanha;Mahin Zohdi Seaf;Kobra Taheri

  • Affiliations:

  • Department of biophysics and biochemistry, Faculty of Science, Tarbiat Modares University, Tehran, Iran;PNU, Hamadan, Iran;PNU, Hamadan, Iran;PNU, Hamadan, Iran;PNU, Hamadan, Iran

  • Venue:
  • BEBI'09 Proceedings of the 2nd WSEAS international conference on Biomedical electronics and biomedical informatics
  • Year:
  • 2009

Quantified Score

Hi-index 0.02

Visualization

Abstract

Approximately 90% of stomach cancers are adenocarcinoma that directly distribute from stomach wall to the neighbor tissue of stomach this kind of cancer is more common in persons who are more than 40 and its spread in men is twice as much as women. Unfortunately, this cancer doesn't have any sign until it develops to its advanced level. Generally biopsy, endoscopy, laparoscopy, ultra sonography, CT scan, x-ray radiography and proper clinical test are applied for cancer detection. By advancing artificial neural network technique and due to the difficulty in detection of stomach cancer with clinical and medicinal parameters, we decided to apply ANN for quick detection of stomach cancer. To do so, we use 50 clinical and medicinal parameters taken from 126 person (90 had cancer and 36 was normal as a testifier). We carried out independent sample T-Test with SPSS software for 50 parameters. Regard to the results of this analysis we selected 8 parameters that had lowest sig for ANN analysis (among parameters whose sig was less than 0.05). These parameters are age, anorexia, weight reduction, MCH, MCHC, Na+ reduction, Ca2+ reduction and X-ray radiography. Selected parameters of 126 persons split to three groups with Matlab software: training group (80%), validation group (10%), and test group (10%). Artificial neural network that we designed has three layers, 8 neurons as input, 8 neurons as hidden and 1 neuron as output. Split data are applied for training network with Levenberg-Marquardt learning algorithm. Finally, Performance of learning was 0.056, Regression coefficient between the output of trained network for test data and real results of test data was 0.927 and the area under ROC curve was 0.883. With these results we can conclude that training process was done successfully and accurately.