An intelligence system approach using artificial neural networks to evaluate the quality of treatment planning for nasopharyngeal carcinoma

ISSN 1992-2248 ©2012 Academic Journals

Full Length Research Paper

An intelligence system approach using artificial neural

networks to evaluate the quality of treatment planning

for nasopharyngeal carcinoma

Tsair-Fwu Lee

*

, Pei-Ju Chao

, Chang-Yu Wang

, Wei-Luen Huang

, Chiu-Ching Tuan

,

Mong-Fong Horng

, Jia-Ming Wu

, Shyh-An Yeh

, Fu-Min Fang

and Stephen Wan Leung

1

Medical Physics and Informatics Laboratory, Department of Electronics Engineering, National Kaohsiung University of Applied Sciences, Kaohsiung, Taiwan.

2

Department of Radiation Oncology, Kaohsiung Chang Gung Memorial Hospital and Chang Gung University College of Medicine, Kaohsiung, Taiwan.

3

Department of Electrical Engineering, National Kaohsiung University of Applied Sciences, Kaohsiung, Taiwan.

4_{Department of Radiation Oncology, Yuan’s General Hospital, Kaohsiung City, Taiwan.} 5

Department of Electronics Engineering, National Taipei University of Technology, Taipei Taiwan.

6

Department of Radiation Oncology, E-Da Hospital /I-Shou University, Kaohsiung, Taiwan.

Accepted 30 May, 2012

The quality of the nasopharyngeal carcinoma (NPC) treatment plans evaluation using three types of artificial neural networks (ANNs) are instructed by three different training algorithms. Three ANNs including Elman (ANN-E), feed-forward (ANN-FF), and pattern recognition (ANN-PR) were trained by using

three different models, that is, leave-one-out (Train-loo), random selection (Train-random), and user defined

(Train-user) method. One hundred sets of NPC treatment plans were collected as the input data of the

neural networks. The conformal index (CI) and homogeneity index (HI) were used as the characteristic values and also to train the neurons. Four grades (A, B, C, and D) were classified in degrading order. The over-training issue is considered between the train data and the number of neurons. The receiver operating characteristic (ROC) curves were obtained to evaluate the performed accuracies. The optimal numbers of neurons for ANN-E, ANN-FF, and ANN-PR, in the loo method are 6, 24, and 9; in the

random-selection method, they are 26, 22, and 4; and in the user-defined method they are 12, 8, and 11 neurons, respectively. The optimal size of train data is 92% of total inputs in the cases of ANN-E and ANN-FF and

76% in the case of ANN-PR. The networks with higher accuracy are ANN-PR-loo (93.65 ± 3.60%), ANN-FF-loo

(88.05 ± 5.84%), and ANN-E-loo (87.55 ± 5.86%), respectively. The networks with shorter training time are

ANN-PR-random (0.55 ± 0.11 s), ANN-PR-user (0.59 ± 0.08 s), and ANN-PR-user (1.07 ± 0.16 s), respectively. The

ROC curves show that the ANN-PR-loo approach has the highest sensitivity, which is 99%. ANN-PR-loo

reduces the amount of trail-and-error during the iterative process of generating inverse treatment plans. It is concluded that the ANN-PR-loo is an excellent model among the three for classifying the quality of

treatment plans for NPC.

Key words: Artificial neural networks (ANNs), dose-volume histogram (DVH), intelligence system,

nasopharyngeal carcinoma (NPC).

INTRODUCTION

Clinically, intensity modulated radiation therapy (IMRT) is the most common technique to deliver radiation doses to nasopharyngeal carcinoma (NPC) patients, because IMRT is capable of delivering a high dose to the irregular tumors, and prevents organs at risk (OARs) and normal tissues from being exposed to radiation. However, it is

usually difficult to complete a suitable IMRT plan at one time because both the patient’s condition and some complex formulas need to be considered simultaneously. The IMRT technique greatly benefits NPC patients, offering much higher treatment quality. The IMRT technique combines several different radiation fields to

produce steep dose-volume histogram (DVH) and isodose curves for the planned target. These steep curves mean that the dose gradient at the border between cancerous and normal tissue varies rapidly. Usually the acceptable dose distribution can be produced by using seven to nine fields (Lee et al., 2008; Oldham et al., 2008).

A treatment plan that results in a higher planning target volume (PTV) coverage and reduce the complications in normal tissues is preferred, and this may be done after several attempts using trial and error. The inverse calculation is one kind of algorithm that is embedded in treatment planning systems (TPS). It is generally used in the optimization procedure for IMRT. The inverse calculation adopts iterative operation and an optimal algorithm to produce varied intensity of treatment beams. This allows IMRT to find a dose that compromises between the PTV and critical organs (Webb, 2004; Leung et al., 2007). The interactive interface is also supported in modern planning systems. The dose-volume based weighting and the priority of the critical organs can be set. Therefore, planners can define some limits for PTV and OARs, which is called constraint-based optimization.

In order to find the solution during the optimization procedure, three steps are performed: (1) determine the constraints and priority setting making up an objective function by the planner, (2) work out the objective function, and (3) evaluate the quality of the treatment plan with the prescribed dose and criteria. These three steps are executed sequentially or iteratively until an optimal solution is reached (Stieler et al., 2009). However, the quality of a final plan depends on the planners’ experiences, which may be learned from others’ experience or published journals (Deasy et al., 2007; Wilkens et al., 2007). It is very time-consuming for a planner to fine-tune for individual optimal solutions. Technically the final result obtained is usually not an optimal solution, but a sub-optimal one. Generally, if we want to find an optimal solution, we have to consider not only a minimum objective function, but also the individual clinical conditions and many parameters not included in the objective function. If an expert knowledge-based system is applied to learn and to accumulate those experiences, then the time taken to create an optimal treatment plan will be reduced. This knowledge-based system is especially effective for complex treatment plans, such as NPC plans.

Artificial neural networks (ANNs) are widely used in the modern sciences (Wu et al., 2009; Bahi et al., 2006;

*Corresponding author. E-mail: tflee@kuas.edu.tw or lwan@ms36.hinet.net.

#Both author contribution equally. Part of this study was presented at the International Conference on Agricultural, Food and Biological Engineering (ICAFBE 2012)

Lee et al. 2077

Vasseur et al., 2008). There are some major researches on ANN models that were used to predict the side effects after radiation therapy. For instant, the leave-one-out (loo), random-selection, and user-defined methods were applied to train the feed-forward ANN model introduced by Su et al. (2005) which is used to predict the probability of pneumonitis after treatment. The sensitivity obtained by the three different methods are 0.95, 0.57, and 0.71 and their respective accuracies are 0.94, 0.88, and 0.90. The ANN is also used to calculate the probability of developing radiation pneumonitis, as proposed by Chen et al. (2007). Based on Chen’s model, the receiver operating characteristic (ROC) curves show that the sensitivity is 0.67, the specificity is 0.69, p = 0.020. Obviously, involvement of the aforementioned non-dose characteristics makes ANNs more generalized. Moreover, Mathieu et al. (2005a) adopts an ANN to optimize the dose distribution. It is applied in treatment plans to make sure the time taken for calculation is acceptable and the error is less than 2%. Isaksson et al. (2005) also uses the feed-forward network to predict the motion of a tumor in the lung during radiation therapy. Results show that this method is better than the conventional one and the self-adaptive filter. Many kinds of treatment techniques introduced by Bortfeld and Webb (2009) are used to reduce the treatment time effectively. Our preliminary result (Chao et al., 2010) shows that a back-propagation model using dose parameters and dose indices can produce high accuracy in evaluating NPC plans. Some applications of ANNs were used in the past to implement 3DCRT plans effectively and efficiently, and a few of them were applied in the expert system of IMRT.

Whether a treatment plan is acceptable or unaccept-able, it usually depends on the planner’s experience. It is time-consuming to evaluate the calculated results and fine-tune the weightings by trial and error. In this study, three types of ANNs are instructed by three different training algorithms to effectively evaluate the quality of the NPC treatment plans. A better match for ANN and the training algorithm will be chosen and established. We aim to help to make an intelligence judgment that reduces the amount of interaction between planner and TPS during the iterative process of generating inverse treatment plans. It can decide whether a plan is acceptable and ranking the quality of treatment plans automatically, therefore, providing an improvement suggestion when the plan was not acceptable.

MATERIALS AND METHODS

Three different neural network models, namely, the Elman (ANN-E),

feed-forward (ANN-FF), and pattern recognition (ANN-PR) models,

are adopted. Each model is worked with three selection methods, named the leave-one-out (loo), random-selection, and user-defined methods, to train the neurons. The overall system flowchart is as shown in Figure 1. The data of DVHs are imported into the untrained ANNs and the training method is selected. We want to

Treatment

planing

Manual

evaluation

ANN

Evaluation

system

Check

criteria

Ready for

delivery

Figure 1. System flowchart for plan quality evaluation and improvement suggestions; ANN: artificial neural networks.

find an ANN model that matches a specific training method to produce the highest accuracy by consideration a mong the conditions of the training time, the number of neurons, the size of training data population, and ROC curves (Chen et al., 2007; Mathieu et al., 2005b). Then, the model can be taken as the best one to evaluate the treatment plans. The basic neural networks structure of the selection procedure is as shown in Figure 2. In the following, parameters and neural networks used are described. Input parameters

According to the International Commission on Radiation Units and Measurements (ICRU) Report 62, the planning organ-at-risk volumes (PRVs) were defined as a safety margin around the OARs, particularly for a high-dose gradient area. In this study, the PRV of the spinal cord was determined by adding a 3D margin of at least 5 mm to the delineated spinal cord. The PRVs of the brain stem and chiasm were defined through addition of a 3D margin of at least 1 mm around the delineated structures. According to the suggestions of Radiation Therapy Oncology Group (RTOG) 0225 (Lee et al., 2003), one hundred NPC samples (NP = N100) are collected as the

inputs, where NP denotes the dimension of sample space and the

suffix p denotes the number of samples in that group. This study was approved by the institutional review boards of the hospitals involved (IRB 99-1420B). Eventually, all the samples will be separated into four ranking classes, named A, B, C, and D. Each class is described as follows:

A: the treatment plan is accepted by physicians.

B: the prescription dose calculated on parallel organs exceeds the criteria.

C: the prescription dose calculated on serial organs exceeds the

criteria.

D: the coverage of PTV does not meet the criteria.

The energy selected is of a 6 MV photon beam and a seven-field IMRT plan is created (Chao et al., 2010). This leads to the reduction of treatment time and enhance the biological effect. Each NPC plan has its own dosimetric indices and parameters (Lee et al., 2010, Lee et al., 2011, Fang et al., 2010; Widesott et al., 2008; Leung et al., 2007), which are discussed subsequently.

Dosimetric parameters

Planning target volume

Three parameters are commonly used to evaluate the coverage of the PTV. V93 means 93% of the total dose is received by 97% of

PTV. The parameter V100 means 100% of the prescription dose

covers more than 95% of PTV. Similarly, V110 means that 110% of

the dose covers less than or equal to 20% of the PTV.

Constraints for the organs at risk

1) Spinal cord (SC): The maximum dose ≤ 45 Gy or 1 cc of PRV ≤ 50 Gy;

2) Brain stem (BS): The maximum dose ≤ 54 Gy or 1% of PRV ≤ 60 Gy;

3) Chiasm: The maximum dose ≤ 54 Gy or maximum dose of PRV ≤ 60 Gy;

Lee et al. 2079

Data Selection Method for Training Phase

Network Model Input

Output

Patients Data

ANN

ANN

ANN

Comparison

Output

Comparison

Train

Train

Train

Figure 2. Basic neural networks structure. ANN: artificial neural networks; ANN-E: the

Elman network; ANN-FF: the feed-forward network; ANN-PR: a pattern recognition network;

Train-loo: ANN with leave-one-out method for training data selection; Train-random: ANN with

random selection method for training data selection; Train-user: ANN with user-defined

method for training data selection.

5) Lens: The maximum dose must be ≤ 10 Gy and as low as possible;

6) Eyes: the maximum dose must be ≤ 45 Gy;

7) Mandible: The maximum dose must be ≤ 70 Gy or 1 cc of PRV and cannot exceed 75 Gy;

8) Oral cavity excluding PTV: the mean dose must be ≤ 40 Gy; 9) Healthy tissue: the mean dose must be ≤ 30 Gy or no more than 1% or 1 cc of the tissue outside the PTV will receive ≥ 110% of the dose prescribed to the PTV.

Dosimetric indices

Conformal index (CI)

This is used to estimate the coverage of PTV (Feuvret et al., 2006).

2 PV TV PTV TV V V CI  

where V_TV is the treatment volume of prescribed isodose lines, PTV

V is the volume of PTV, and TV_PV is the volume of V_PTV

within V_TV. The best conformal case is the value of CI equal to 1.

Homogeneity index (HI)

This index describes how the homogeneity varies within the PTV.

% 95 % 5 D D HI

where D_5% and D_95%are the minimum doses delivered to 5 and 95% of the PTV. A higher HI indicates poorer homogeneity.

Therefore, there are 14 indices included in D = [V93, V100, V110,

SC, BS, rt Parotid, lt Parotid, Lens, rt Eye, lt Eye, Oral, Mandible, CI, HI] which are presented in this paper as the input vector (rt: right side, lt: left side).

Training parameters

Three distinct training methods are considered separately as follows:

Conclusions

ANN-PR-loo reduces the amount of trail-and-error during

the iterative process of generating inverse treatment plans. It is concluded that the ANN-PR-loo is an excellent

model among the three for classifying the quality of treatment plans for NPC. This system is able to classify the calculated result and offer suggestions to planners that reduce the amount of interaction between planner and TPS during the iterative process of generating inverse treatment plans. It is a convenient and effective way to evaluate the quality of treatment plans.

ACKNOWLEDGEMENTS

This study was supported financially, in part, by grants from Kaohsiung Chang Gung Memorial Hospital (CGMH) and the National Science Council (NSC) of the Executive Yuan of the Republic of China. “CMRPG890062”, “NSC 100-2221-E-151-003.

REFERENCES

Bahi J, Contassot-Vivier S, Makovicka L, Martin E, Sauget M (2006). Neural network based algorithm for radiation dose evaluation in heterogeneous environments. Artif. Neural Networks–ICANN, pp. 777-787.

Bassi PF, Sacco E, De Marco V, Aragona M, Volpe A (2007). Prognostic Accuracy of an Artificial Neural Network in Patients Undergoing Radical Cystectomy for Bladder Cancer: a Comparison with Logistic Regression Analysis. BJU int., 99(5): 1007-1012.

Bortfeld T, Webb S (2009). Single-Arc IMRT? Phys. Med. Biol., 54(1): N9-20.

Chao PJ, Lee TF, Huang WL, Chen LC, Su TJ, Chen WP (2010). Applications of Artificial Neural Network to Building Statistical Models for Qualifying and Indexing Radiation Treatment Plans. World Acad. Sci. Eng. Technol., 70: 37-42.

Chen S, Zhou S, Zhang J, Yin FF, Marks LB, Das SK (2007). A Neural Network Model to Predict Lung Radiation-Induced Pneumonitis. Med. Phys., 34(9): 3420-7.

Cheng Y, Qi WM, Cai W (2002). Dynamic properties of Elman and modified Elman neural network. in IEEE, 2: 637-640.

Deasy JO, Alaly JR, Zakaryan K (2007). Obstacles and Advances in Intensity-Modulated Radiation Therapy Treatment Planning. Front. Radiat. Ther. Oncol., 40: 42-58.

Duin R, Juszczak P, Paclik P, Pekalska E, De Ridder D, Tax D, Verzakov S (2007). A matlab toolbox for pattern recognition. Delft University of Technology, PRTools4. 1.

Fang F, Tsai W, Lee T, Liao K, Chen H, Hsu H (2010). Multivariate analysis of quality of life outcome for nasopharyngeal carcinoma patients after treatment. Radiother. Oncol., 97(2): 263-269.

Feuvret L, Noel G, Mazeron JJ, Bey P (2006). Conformity index: A review. Int. J. Radiat. Oncol. Biol. Phys., 64(2): 333-342.

Gulliford SL, Webb S, Rowbottom CG, Corne DW, Dearnaley DP (2004). Use of artificial neural networks to predict biological outcomes for patients receiving radical radiotherapy of the prostate. Radiother. Oncol., 71(1): 3-12.

Isaksson M, Jalden J, Murphy MJ (2005). On Using an Adaptive Neural Network to Predict Lung Tumor Motion During Respiration for Radiotherapy Applications. Med. Phys., 32(12): 3801-9.

Klopf AH, Gose E (1969). An Evolutionary Pattern Recognition Network. Systems Science and Cybernetics, IEEE Trans., 5(3): 247-250. Kolasa M, Wojtyna R, D ugosz R and Jo wicki W (2009). Application of

Artificial Neural Network to Predict Survival Time for Patients with Bladder Cancer. Comput. Med. Act., pp. 113-122.

Lampariello F and Sciandrone M (2001). Efficient Training of RBF Neural Networks for Pattern Recognition. IEEE Trans Neural Netw., 12(5): 1235-1242.

Lee JSJ, Bezdek JC (1988). A Feature Projection Based Adaptive Pattern Recognition Network. in Neural Networks, 1988. IEEE Int. Conf., 1: 497-505.

Lee N, Kramer A, Xia P (2003) 'RTOG 0225: a phase II study of intensity modulated radiation therapy (IMRT)+/-chemotherapy for nasopharyngeal cancer'.

Lee TF, Chao PJ, Fang FM, Su TJ, Leung SW, Hsu HC (2010). Helical Tomotherapy for Single and Multiple Liver Tumours. Radiather. Oncol., 5: 58.

Lee TF, Chao PJ, Wang CY, Lan JH, Huang YJ, Hsu HC, Sung CC, Su TJ, Lian SL and Fang FM (2011). Dosimetric Comparison of Helical Tomotherapy and Dynamic Conformal arc Therapy in Stereotactic Radiosurgery for Vestibular Schwannomas. Med. Dosim., 36(1): 62-70.

Lee TF, Fang FM, Chao PJ, Su TJ, Wang LK, Leung SW (2008). Dosimetric Comparisons of Helical Tomotherapy and Step-And-Shoot Intensity-Modulated Radiotherapy in Nasopharyngeal Carcinoma. Radiother. Oncol., 89(1): 89-96.

Leung LHT, Kan MWK, Cheng ACK, Wong WKH, Yau CC (2007). A new dose–volume-based Plan Quality Index for IMRT plan comparison. Radiother. Oncol., 85(3): 407-417.

Liu H, Wang S, Ouyang P (2006). Fault Diagnosis Based on Improved Elman Neural Network for a Hydraulic Servo System. in Robotics, Automation and Mechatronics, 2006 IEEE Conf., pp. 1-6.

Luo J, Yu G, Yang H, Zhao H, Ai B (2005). Bus Protection ANN Model with Function Approximation Ability. in Power Engineering Conference, 2005. IPEC 2005. The 7th Int., pp. 1-478.

Maruno S, Imagawa T, Kohda T, Shimeki Y (1993). Object Recognition System Using Temporal Pattern Recognition Networks with Quantizer Neuron Chip. in Neural Networks, 1993. IJCNN '93-Nagoya. Proceed. 1993 Int. Joint Conf., 2: 1285-1288.

Mathieu R, Martin E, Gschwind R, Makovicka L, Contassot-Vivier S, Bahi J (2005a). Calculations of Dose Distributions Using a Neural Network Model. Phys. Med. Biol., 50(5): 1019-28.

Mathieu R, Martin E, Gschwind R, Makovicka L, Contassot-Vivier S, Bahi J (2005b). Calculations of dose distributions using a neural network model. Phys. Med. Biol., 50: 1019.

Oldham M, Sakhalkar H, Guo P, Adamovics J (2008). An investigation of the accuracy of an IMRT dose distribution using two-and three-dimensional dosimetry techniques. Med. Phys., 35: 2072.

Qi Z, Liu M, Wang H(2008). Process Modeling Method Based on an Improved Elman Neural Network. in IEEE, pp. 8188-8192.

Stieler F, Yan H, Lohr F, Wenz F, Yin F-F (2009). Development of a neuro-fuzzy technique for automated parameter optimization of inverse treatment planning. Radiather. Oncol., 4(1): 39.

Su G, Wang L, Xu Y, Wang X (2007). Date Gas Load Forecasting With OIF-Elman Network. in Control Conf. 2007. CCC 2007. Chinese, pp.167-170.

Su M, Miften M, Whiddon C, Sun X, Light K and Marks L (2005). An Artificial Neural Network for Predicting the Incidence of Radiation Pneumonitis. Med. Phys., 32(2): 318-25.

Ulug ME (1996). Combining Fuzzy Pattern Recognition and Fuzzy Control in an AI Driven Neural Network. in Fuzzy Systems, 1996. Proc. Fifth IEEE Int. Conf., 3: 2071-2077.

Vasseur A, Makovicka L, Martin E, Sauget M, Contassot-Vivier S, Bahi J (2008). Dose calculations using artificial neural networks: A feasibility study for photon beams. Nuclear Instruments and Methods in Physics Research Section B: Beam Interact. Mater. Atoms., 266(7): 1085-1093.

Weaver CS (1975). Some Properties of Threshold Logic Unit Pattern Recognition Networks. IEEE Trans. Comp., 24(3): 290-298.

Webb S (2004). Contemporary IMRT: Developing Physics and Clinical Implementation. ed., Taylor & Francis.

Widesott L, Pierelli A, Fiorino C, Dell'Oca I, Broggi S, Cattaneo GM, Di Muzio N, Fazio F, Calandrino R, Schwarz M (2008). Intensity-modulated proton therapy versus helical tomotherapy in nasopharynx cancer: planning comparison and NTCP evaluation. Int. J. Radiat. Oncol. Biol. Phys., 72(2): 589-596.

2088 Sci. Res. Essays

Treatment Planning Based on Prioritizing Prescription Goals. Phys . Med. Biol., 52(6): 1675-92.

Wu HY, Hsu CY, Lee TF, Fang FM (2009). Improved SVM and ANN in incipient fault diagnosis of power transformers using clonal selection algorithms. Int. J. Innov. Comput. Inf. Control, 5(7): 1959-1974. Yuan-Chu C, Wei-Min Q, Jie Z (2008). A new Elman Neural Network

and its Dynamic Properties. Cybern. Intell. Syst. 2008 IEEE Conf., pp. 971-975.

Zhang LD, Jia L, Zhu WX (2010). Overview of Traffic Flow Hybrid ANN Forecasting Algorithm Study. Comput. Appl. Syst. Model. (ICCASM), 2010 Int. Conf., 1: 615-619.